I procrastinated fixing Familiar’s onboarding way too long. So many insights built up from watching people, and I didn’t know where to start, so I wanted Opus 4.6’s help. I brain dumped a really messy doc: the kind I would never give to a human teammate, but that an enthusiastic, unfocused CEO might slack me on a Saturday night (we’ll come back to this analogy).

You’ve heard the next part a million times: Opus did a fantastic job blah blah blah, and it even created the issues in Linear, prioritized, with tight descriptions, organized by neat milestones. Agentic agentic insane insane.

But enough about AI, let’s talk about me.

I got to be a cross between a caffeinated product designer and a salesperson with “tons of ideas for the product.” That was WAY more fun than PMing! Meanwhile, AI structured my thoughts, reminded me of the strategy, cut scope, and broke things into concrete phases. (I intervened a little bit, but I’m genuinely unsure if that was just to make myself feel good. Also it’s March 2026, and this is the worst it’ll ever be.)

What’s my added value? Not much. I just happen to have access to context that it doesn’t. I’m a gatekeeper, and that won’t last.

You know what that reminds me of? The best product teams I’ve been a part of. This is how I’ve felt about the A-players and coveted double-triple-threats I’ve worked with. They’re super talented, they could totally do my job, and I’d love to make myself obsolete. Long before AI, the best version of my job was to sit next to experts and builders, connect them to as much context as possible (customer and company) and get out of the way.

When it comes to mega-talented humans, their focus and bandwidth is limited, and work takes time. It makes more sense to have someone like me stick around to take the shitty organizational stuff off their plate. But that’s like, it.

So, AI is reviving my all-time favorite feeling at work:

1. Whoa, this can take my job.

2. Woo!!

3. How can I make myself obsolete even faster?

I’m giving away my Legos… to AI

Tyler Palmer, Patreon’s first employee and SVP of operations for 10 years, once told me that Patreon felt like a new company every 6 months. I’d say that describes our entire industry now.

Instead of a chaotic, hypergrowth company, we’re all in a chaotic, hypergrowth industry. I need you with me on this, because it’s the premise for everything else I’m going to say.

In 2015, Molly Graham wrote an article for First Round Review about the “ambiguity, chaos and stress that comes with doubling or tripling your team every six months.” It’s called “Give Away Your Legos” and if you work in startups you’ve already read it.

Read this and tell me she’s not talking about AI, too:

“As you add people, you go through this roller coaster of, ‘Wait, is that new person taking my job? What if they don’t do it the right way? What if they’re better than me at it? What do I do now?’” says Graham. “These are some strong emotions, and even if they’re predictable, they can be unnerving.”

And how about this:

“There’s a lot of natural anxiety and insecurity that the new person won’t build your Lego tower in the right way, or that they’ll get to take all the fun or important Legos, or that if they take over the part of the Lego tower you were building, then there won’t be any Legos left for you. But at a scaling company, giving away responsibility — giving away the part of the Lego tower you started building — is the only way to move on to building bigger and better things.”

Molly wrote this 11 years ago. Is this not the best advice anyone’s given about being a knowledge worker in 2026?

If Claude is the PM, what are my new Legos?

As Claude PMed me, I felt my role elevate. I could be the people whose chaos my job was to turn into order.

I could [try to] be that inarticulate, absent-minded, yet visionary founder. I could be the customer-facing person who’s full of visceral insights. I could be a designer immersed in inspiration, imagination, and creating 1-year vision flows, detached from bug reports, tech debt, and sprint planning.

Could I just spend all my day outside the building + sitting in product reviews?

Claude could handle strategy, opportunity assessments, roadmap prioritization, scoping, phasing, de-risking, measurement strategy, managing stakeholders, rollouts…

There’s an emotion coming up for me right now, but Molly called it 11 years ago:

“A week ago, someone might have told you they hate their Legos and want to get rid of them. But as soon as you hire someone else, they suddenly want to hang on to all of them.”

What if judgment, product sense, order of operations and scoping are actually low-level mental operations? The kind that Microsoft Excel abstracted away for hard skills that used to be societally admirable, and today are considered menial.

Back to Molly:

“That’s one of the other counterintuitive things: Adding people doesn’t mean there’s less work for the people that are already there. It means that the entire company can do more…

People think, ‘Oh, that person joined! Now I can finally work a little less.’ But that’s almost never what happens,” Graham says. “Adding people is the opportunity to find a new job (or the new version of your old job).”

I love this. AI taking over PM work doesn’t mean “people-who-previously-wore-PM-hats” have less to do. It means we can do more—if we’re willing to do a different job.

So, maybe my Legos now are to set up AI for success. The next incarnation of my job is to do for AI what great executives do for their product builders:

• Provide missing context

• …and build systems so it can get that context without me

• Give clarity on strategy, target customer, constraints

• Own hard, fateful decisions

• Keep things focused, and with healthy room for chaos

• Know when to micromanage and when to empower

• Separate signal from noise

• Bring values, conviction, mission, and optimism

And no, that’s not PMs pushing features to production. It’s unimaginative to take senior PMs and turn them into junior engineers. Why give away your Legos to simply take somebody else’s?

Remember when Mixpanel and Amplitude got widely adopted in the 2010s? Remember what happened to the role of product data analyst? Yes, companies hired fewer analysts, and later. But each one of them had way more interesting work, could support more teams, and engage at a way more strategic level. Self-serve analytics tools elevated the analyst role.

The tools also changed hiring. Because each analyst role was so high leverage (and high opportunity cost), only the most strategic-minded, technically skilled data analysts were worth hiring. Better tools raised the bar for excellence.

That’s where today’s product managers are heading. Each “former PM” is going to be so high leverage that the opportunity cost of a wrong hire is even bigger and bigger. But wasn’t that always the trend?

I think that’s great news. With so much opportunity, if we have enough excellent people, they can all get hired, because they’re just worth it. Remember: great companies want to be able to do more with more, not the same with less.

What do I do right now? (IC PM edition)

Intuitively, our brains are asking “How do I hold on to my job?” The more effective phrasing is “how can I be as valuable as possible.”

1. Stay in the weeds and attuned to your customers, team, and company - the way a good founder or executive might

2. Stay hands-on with AI tools (just use it, it might suck or might be awesome, just keep using it regularly so you can notice that)

3. Notice little pockets of opportunity where you can set it up for success

4. Help your entire team with this

5. Help your entire org with this

Just like when leading excellent teammates, recognize opportunities to be valuable. And when those moments come, give away your Legos.

What do I do right now? (Leadership edition)

1. Do the above, publicly, as a personal example.

2. Give your reports room to do the above.

3. Be flexible on role expectations, and value the people doing things differently

Aside from that, I’ll hand the mic to Molly:

“Honestly, the best thing you can do is normalize what people are experiencing. As a leader, you want to head it off at the pass and proactively say, ‘Hey, this is what you can expect to feel during this time of growth. It’s pretty universal. Other people are going through the same thing. I’ve been through it too. There’s no need to be scared.’”

And then you can accelerate it:

“That’s when our boss did this brilliant thing — he gave her a huge goal…. Immediately, my co-worker let go of every single Lego she was holding and ran to the new project because she was so scared and excited. It was like someone had flipped a switch and she was suddenly like, ‘Good luck, peace out!’”

If your reports are surfers, you’re the surf instructor giving their boards a little push down the wave.

We all now work in a hypergrowth situation

It’s exhausting. Change and chaos mean bringing your best energy to your job. That’s really hard to do without a break, and I don’t want to romanticize anything. If you're between jobs right now, hypergrowth probably sounds like a cruel word.

The good part of hypergrowth (AI-driven or not) is it’s a chance to grow a ton. Under what other circumstances can you regularly elevate your role?

I believe that we all have the ability to galvanize people and to step into chaos, to learn things without being taught, to create clarity for others and operate with little structure… whether or not we’ve brought that to work.

Stay in the game, stay hands-on, and you’ll notice opportunities to make yourself obsolete. When the opportunity comes, give away your Legos.

Claude would think, ask us a question, and we’d go off and discuss. Then we’d dictate our answer back.

“How time-sensitive is the notification?”

“How many results might pile up before someone checks?”

“What if we had no notification?”

In between Claude’s questions, we’d have a lengthy conversation. We didn’t transcribe everything we said, just the bottom line. When it felt like one of us was about to land an important point, we’d hold down the transcription button, and hit enter.

Claude got our creative juices flowing. It kept us moving, and got us to think bigger and articulate a framework for our decision, faster. Claude was an excellent addition to the meeting. (Not everything Claude said was brilliant, but neither are most things I say.)

I realized this wasn’t my first time pulling AI into a live meeting. When Laurel was traveling and we had to make some complicated decisions, we did a Zoom call. We really wanted Claude’s help, so:

1. I screenshared Claude

2. I turned on Granola

3. We talked about it

4. We paused, pasted our transcript into Claude, and read the answer

5. We talked about it

6. We paused, pasted our transcript into Claude, and read the answer

7. etc.

Full disclosure: this was super manual. Each time I had to paste the transcript into TextEdit so I could truncate just the transcript from where we left off. But it worked.

Another option is to have the call on your laptop speakers (not headphones), and hold down the dictation button (I use Wispr Flow these days).

This also works one-sided: a friend in sales told me about a call that was going south. Mid-conversation he copied his Granola transcript into AI, that gave him one discovery question to ask. It changed the tone of the entire conversation, and he eventually closed the deal.

Imagine how valuable this could be:

• Live coaching during a sales call

• Live coaching during a hard conversation

• Two people having a hard conversation and both want AI present (like couples therapy for the smaller things where it doesn’t make cost benefit to hire someone)

I’d love for this to be less clunky:

Granola, Grain, Fathom, etc. folks: can you make the live transcript available through your MCP integration? Like live_conversation_latest_since(timestamp) or something, so AI can chime in and has full conversation context. (Or leave the tool response hanging until a natural pause in the conversation, and then return the incremental transcript.)

And you, AI voice mode providers: what if AI could just accept a Zoom/Meet/Teams invite link and join in?

I’m hoping these exist and I’m late to the party. If you personally use one, reply and let me know!

Maxim and I became Claude Code’s subagents

Think about what was actually happening in that brainstorm:

1. Claude would ask us a question (aka prompt us)

2. Maxim and I would talk for a while (aka reasoning)

3. Then we’d distill the bottom line and dictate it back to Claude (aka tool response)

4. Claude would process it, prompt us with a new question (main agent orchestrating us)

Sounds a lot to me like we were AI’s subagents. I don’t mean this in a philosophical Matrix sense: this exercise deepened my understanding of subagents, by noticing I was acting as one.

Remember, a “subagent” is when one AI agent (aka chat thread) opens a blank, fresh AI agent (aka chat thread), to get an answer without wasting its own context window tokens. The subagent gets a request (a prompt, written by the original agent) and gives back a response (the bottom line answer, returned to the original agent). If you want to sound smart in a meeting, the fancy term for this is “uncorrelated context window.”

In our brainstorm, Claude’s prompt to us was its multiple choice questions (e.g. “how time-sensitive is the notification?”) Our subagent response was: “Not urgent, but important. Can have a delay of a few hours but not more” Claude processed that and called us again with a new prompt.

Btw, what I love about Cursor is you can see this very easily. When a model calls a subagent, you can click into it to see the original prompt, actions, and response to the main agent (another way to see this in action was I had Claude Code write and workshop standup comedy).

“Agent swarms” are just subagents with friends

Let’s throw a wet blanket on the buzzword: a “swarm” is just multiple subagents running at the same time.

I haven’t encountered a killer use case for AI swarms. As with humans, throwing more parallel agents at a problem feels limited. As Fred Brooks put it, "The bearing of a child takes nine months, no matter how many women are assigned."

Even if swarms are incremental to subagents, they’re still a cool idea.

Imagine for a moment if the “AI swarm agents” were just humans. What would that look like?

• AI sends out a batch of surveys

• AI facilitates a large-scale brainstorm, cross-pollinates ideas, synthesizes the results

• A 360 review orchestrated by AI

• AI collecting product feedback at scale

In that sense, are UserTesting, Mechanical Turk, and SurveyMonkey the original “swarms”? It’s a fun thought, and makes the buzzword more accessible.

Taking the “sub” out of subagents

As a subagent, I resent the implication of hierarchy. Subagents aren’t really “sub.” They’re just participants in a conversation, just like users and tools.

Maxim and I brought connection to the real world, a lifetime of context, taste, and judgment. Claude brought knowledge and extra creative brain cells. We all read the messages, thought about them, and gave our best response.

When I pull AI into my real work, the buzzwords lose their bite, and my AI product sense gets a little stronger. This started as pulling AI into a brainstorm about notifications. Turns out it was also the clearest explanation of subagents I’ve found: just being one.

Quick Familiar update: You all blew our expectations away. We’ve run 52 onboarding conversations and made a bunch of improvements thanks to all of you. We weren’t able to schedule with everybody who wanted to, and our bottleneck was creating enough time to balance learning with acting on those learnings.

If you reached out and we haven’t connected yet, we saved every message and you’ll be the first people we schedule with. For now, go ahead and download it so it’s building up context. When we do meet, we can dig into some really interesting use cases together.

Our first goal was making single-player mode really great, and we’re now researching multiplayer mode (i.e. what happens when a critical mass of people in your organization install Familiar and compound each other’s use of AI). If you lead a team and that feels exciting, we’d love to hear why - just hit reply.

It took me a long time to vibe code a landing page for Familiar

We intentionally didn’t build a landing page for Familiar when we launched it three weeks ago. We wanted everyone who signed up to schedule 10 minutes with us for onboarding so we could learn as much as possible. After 52 conversations, it was time to make it easier for people to just download and install it.

Maxim is the actual talent, so I took the task of making a landing page.

I wanted to share this project specifically because it’s not technically complicated. It’s a static landing page. I chose the least sexy, most trivial-sounding thing to show you how much work and research and customer knowledge still went into getting something I’m proud of. And how much I still had to partner with AI to avoid slop.

I’m worried there’s a weird incentive in our industry right now to show how quick things were and how easy things were for you, as a mark of being AI-forward. We say things like “just vibe code it” like that’s some kind of answer. So I wanted to share my experience, theater-free.

I did this in two phases: getting the content right, and vibe-designing-coding the actual page. Spoiler: The hardest part was going to be figuring out what should actually be on the page. The second hardest thing would be designing something that looks great.

If you’re feeling impatient and want to see the result: https://www.looksfamiliar.org/

Phase 1: using Cursor to figure out the content

There was no need for AI to slop-generate this content. We had a ton of raw customer insights to work with and thus a chance to be really thoughtful. In this case, “fast” or “efficient” was not the goal, until I nailed “effective.” And that bottleneck isn’t coding, it’s how fast we could learn from our users.

I started by creating an empty markdown file in my project as a spec for the landing page, and filled it out with my initial thoughts.

Turning user conversations into a super-precise FAQ

A lot of the same questions would come up often, and the landing page was the right place to address a lot of them. We had the luxury of being really precise about this because we knew exactly how many people asked which question.

Here’s the prompt I gave Cursor:

We've done a lot of onboarding conversations with our first few dozen
users and I'd like to really get a sense for what are the common
questions that people asked and that we had to answer. Let's be really
comprehensive and create a superset here. Come up with a ranked list
of the main ones that come up all the way down to the less common
ones, and really capture them all inside of the section on line 9.

I used plan mode so it would ask me questions before diving in, to align before it runs off. (Plan mode is the same as just telling it to ask you questions and then save a plan somewhere, but it’s a nice UI affordance).

I’m using Claude Opus 4.6, and I noticed it’s spawning subagents (Cursor’s faster, cheaper Composer model) to do the actual research. Opus acts like an executive - it doesn’t use a ton of context itself. Most of the context gets consumed by these subagent threads, and only the bottom line comes back to the main thread.

I love that you can click into each subagent and see the prompt that Opus wrote, what the subagent thought, and what came back. This way, the agent can cover a massive amount of ground with very few tokens.

Adding the answers

Once I had the ranked list of questions, I asked it to add the answers. Again it used subagents, and I liked watching how it divided the work. It wasn’t splitting by question - it was splitting by person. One subagent handled one group of transcripts, another handled a different group. Same questions, different sources.

Now I had a juicy spec with use cases, FAQs, and real answers from real conversations.

Getting a second opinion

I wanted AI’s opinion on the structure before filling in the details

Please take a look at [the spec] and given all the conversations in
[the transcripts folder], what would make this a really effective
landing page for our target audience? What's missing? I would love
your help as a partner on this.

Even though the task was pretty well defined, it went wide. It read our original blog post. It read the current placeholder draft on looksfamiliar.org. I didn’t ask it to do any of that. I guess you’d call that “high agency” (I’m relieved to have a word for this other than “agentic”).

Its feedback was really, really, really good. Like better than most marketing people I’ve worked with.

I implemented AI’s feedback semi-manually - I wanted to be really familiar with everything it was telling me and apply my own judgment on what belongs where. Funny how the bottleneck is me catching up and staying in context.

Filling out the landing page content, section by section

So at this point I had a doc with a bunch of new sections. I asked AI to help fill each one with the best material we’ve gathered so far: transcripts, blog post, whatever was relevant.

I could see I was on a path to lots of slop, so I stopped it and added the following to my chat:

Let's do this one thing at a time. Each section and your feedback.
Let's go one piece of feedback at a time. Find the section I created
for it. Ask me anything, one question at a time.

“One question at a time” pulls stuff out of me that I wouldn’t have articulated upfront, and it gives me chances to apply my taste along the way.

Whittling down to a lightweight landing page spec

The result was a super set of things that could go on the landing page in a rough structure. I wanted to really curate and whittle this down to something very focused and purposeful that I could give to AI to implement. It would have only what I wanted to appear on the page, and no fluff.

Phase 2: vibe designing and coding the page

My biggest fear was it making something that was slightly off, and then being stuck in vibe-coding mud for days.

To avoid this, I chose a design system (any design system) that helps keep things maintainable and coded quickly. I decided on shadcn which is a basic frontend foundation for lots of great UI libraries (that I can then mix and match).

I decided to switch to Claude Code for this part. Here’s the prompt:

I want you to take the lightweight website spec for the landing page
and turn it into a landing page that's a sibling to the current
marketing website directory, so we can see them side by side. Make it
a static site I can deploy on Vercel, just like the other one.

Treat the markdown file with all the content as an open-ended spec.
Stay very true to that content, but I want you to decide which UI
components and blocks to render them with.

Use ShadCN with ShadCN Marketing blocks.

Stylistically, start inspired by Mobbin - their typography and their
vibe. Beyond that, use the /frontend-design.

Let's go section by section. Tell me what you think, give me a link
to preview what you want to use for each section, render it, show it
to me, and then I'll give you feedback. Then we'll go to the next
section and build the landing page together that way.

I'm looking to you to lead me and bring your creativity here. But
also be a consultant and ask me the important questions you need to
make these decisions. Ask me one question at a time, at any point.

Claude Code’s frontend design skill

Huge credit to Eric Xiao for sanity checking my process, and pointing me to use the /frontend-design skill. Last November, Anthropic published a great design skill. It’s not long, you can read the entire skill’s text here.

This together with a design system makes things that look way better. This is Claude Code using the frontend skill and existing components together. (Btw, this skill can be used in Cursor just as well - it’s just a text file with instructions).

Beyond that, making something great is very much the timeless design process. Together with Aman Khan, Eric covers this in depth on this process and designing with AI in their workshop, Claude Code for PMs, which has a cohort starting in 11 days.

Iterating on the design, section by section

It built the hero section and asked what I thought. I didn’t like the typography - Claude Code picks that same font a lot. Maxim was sitting next to me and agreed it was weird. So I told it what I wanted instead.

This continued section by section. It mostly knocked them out of the park, with only minor babysitting. The tweaks and time invest were mostly for capturing my taste preferences, not “you made an obvious mistake I have to fix.”

Closing feedback loops

One thing that made this work well: Claude Code (in the Claude desktop app) as well as Cursor both take screenshots of what it built and look at them. The agents can close their own feedback loop.

This is one of the most powerful things that’s happened with coding agents. When an agent can close a feedback loop, whether that’s unit tests, API endpoints, logs, or looking at its own screenshots, it becomes much more autonomous.

If you remember the early days of vibe coding, every error you had to copy-paste in. You had to tell it what it did wrong. You had to close the loop for it. Now it closes many of those loops itself.

Learning along the way

There were little bits of cool functionality I couldn’t wrap my head around. So I’d just ask: “by the way, how’d you do that?” Any question that came to mind. I turned those moments into learning opportunities and had it teach me.

I guess that’s another instance where the bottleneck is me keeping up with AI.

Deploying

Let's deploy this to Vercel. It should also appear on looksfamiliar.org.

To make this work, you can either install the Vercel MCP or CLI. They’re both the same idea: an interface for an agent to directly control Vercel and take care of deployments instead of you having to click around on their website.

That was hard work. The best kind.

This is a static landing page. One page. No backend. No database. And it still took me real focus across multiple sessions.

I think as an industry we’ve temporarily regressed. We’ve replaced outcomes with outputs and inputs. We’re valuing how quickly something took over what goal it had and whether it achieved that goal.

This shit is hard, and there are a lot of people with incentives to show you that it’s easy and magical. When you see those demos, it’s worth asking:

• How high are the stakes?

• Are those real things?

• Are they showing the full picture?

• Have those outputs achieved outcomes?

Landing pages are deceptively hard, even with AI

They take customer intuition, customer context, customer conversations, knowledge of storytelling, and knowing what your customers’ objections are.

The effort that went into this landing page started long before the process I wrote about here. It’s all the effort that came before it: we had a prototype (my previous blog post and all your sharp comments) and a validation process (dozens of one-on-one onboarding conversations that we transcribed).

AI is amazing, and I couldn’t completely delegate this

It still required a lot of brain energy and thoughtfulness and attention and focus and blocking off my calendar and getting into flow. I can’t just tell it “go do it” like a VP of marketing who has all the context. But it definitely let me do what I feel is a really good job in a really short amount of time.

Vibe coding a landing page with AI did not feel like a magic genie. I’d describe it best as a pair programmer partner. Like pairing, I came out of this mentally exhausted in the best way possible.

For most strategic tasks, AI remains in the “amazing pairing partner” category. It’s exciting to watch it move into the “teammate I can delegate with my eyes closed.”

Every time I let AI run loose, I got slop. Every time I kept it tight - one section, one question, react, adjust - the output was good.

I’m ridiculously appreciative even for that. The outcome is way better, it still took way less time than ever before. Check out the result: https://www.looksfamiliar.org/

If you got this far and have feedback on what sucks/needs to be clearer/confused you - hit reply. It goes to my personal gmail.

• “What decisions did we actually make this week that aren’t documented anywhere?”

• “What did I even DO today? Where’d that time go?”

• “What two things things this week seem unrelated but might actually connect?”

All three answers were a kick in the pants. The third one was so spot-on, it made me uncomfortable.

How did Opus 4.6 know so much about me? Sure, it’s a smart cookie that has context on my work, memory, and integrations. And now it has one more thing:

Using Familiar, AI now watches my computer screen. All. The. Time.

What is Familiar?

Familiar is a desktop Mac app that takes everything that passes through your screen (and clipboard) and saves it as context for your existing AI (OpenClaw, Claude Cowork, Claude Code, Codex, Cursor, Antigravity, you name it).

Familiar is free, open source, and offline.

Familiar runs in the background from the moment you start your Mac.

You install it, and your AI just gets smarter. It runs quietly in the background from the moment you start your Mac. No habits to build, no interface to learn.

What can I do with it?

I finally have a partner who sees everything I see. Imagine asking your favorite AI agent:

• "I'm going on parental leave. Based on my last 3 months, what does my replacement need to know about each initiative?"

• “I need to write my weekly status update and my brain is completely empty”

• “Who am I not talking to enough?”

• "What were the exact arguments for and against the decision we made about the pricing model? I need to brief my new director."

• “I just had a breakthrough solving a problem. Reconstruct the exact sequence so I can document/communicate it.”

• "I've been bouncing between Slack threads and Google Docs for two hours and my brain is soup. What are the open threads I still need to respond to?"

The answers are good, and they get better with every passing minute.

How does it work?

1. Familiar looks at your screen every few seconds and converts it to text using Apple’s native OCR.

2. It also saves everything that passes through your clipboard (this includes most third-party speech-to-text tools, too)

3. All these text files go in a local folder of your choosing (we recommend placing /familiar/ inside wherever you work with your AI agent)

That’s it. We kept it simple. Just a menu bar icon, so you can pause or quit anytime.

As you use your AI thinking partner, it can use that context to help you. (To make that smooth, Familiar can auto-add a Skill, a text file that tells your AI when and how to use your new folder.)

We’ve observed AI agents use Familiar as one more layer on top of tools, file system, skills, and the LLM’s own knowledge. Context compounds: each source fills the blind spots of the others, and the AI starts connecting dots none of them could surface alone.

We’re excited about screenshots. Screenshots have information that MCPs and APIs don’t: behavior signal on where you spend your focus and attention. The Google documents and Slack threads where you linger speak volumes about what's important. Where you spend your time is valuable context.

Screenshots are also immune to SaaS companies anxiously walling off your data (not naming names).

And yes - OCRing screenshots is messy, often illegible to humans. LLMs are beautifully antifragile to messy input.

We stand on the shoulders of giants: Rewind, Recall, Dayflow, and others who watched your screen before us. We’re especially grateful to Louis Beaumont, creator of Screenpipe, who met us in person and was super encouraging.

Familiar is unique in two ways:

1. We’re focused on one use case: raw context for your existing AI agents.

2. Timing

Why now?

The models are good enough.

The latest AI models are resourceful. You can give them a directory full of files, and watch them glide across stupidly large amounts of messy text with the simplest of tools. They’re increasingly proactive in managing their own context with tactics like context editing and subagents. As a result, Familiar doesn’t need to capture everything perfectly; it’s just the starting clues for an agent.

Something about AI has changed in recent weeks. When I watch my AI agents at work, I feel more like a biologist than an engineer. I spend more time humbled by complexity-bigger-than-my-brain, less time inspecting the gears of a clock.

I’ve stopped patronizing and coddling AI. These days I just give it all the context I have, get out of the way, and let it rise to the occasion.

In other words, the Bitter Lesson has reached our use case. The answers aren’t always perfect, but like everything with AI, I take the good and leave the bad. And the good is pretty damn good.

Tal, this scares the crap out of me

It scares us, too. We’re leaning into it.

This is a good time to introduce my co-founder, Maxim Vovshin. He’s an early contributor to OpenClaw, ex-Orca Security, and ex-military intelligence. (He’s also my second-cousin’s-husband, which in a country of 10 million, could refer to 50% of the population).

We decided that for this crazy idea to work, it would have to be:

1. free

2. open source

3. offline

As a result, Familiar is the security equivalent of “taking a screenshot and saving it to your hard drive.” That’s allowed even in windowless, underground rooms in the Pentagon.

From there, it's your LLM of choice. For sensitive work information, that should be the provider you have an enterprise contract with, or your company’s internally hosted models.

What if I Google something embarrassing?

You can pause or quit Familiar anytime. You can also directly go to the file and delete as much of the context as you want.

Why did you build this?

We want this.

We’re also not alone. I have the privilege of accessing hundreds of product people who are doing everything right when it comes to using AI. Yet in 1-1 conversations, they all hit the bottleneck of “keeping AI updated.”

"I feel like a lot of what I do is getting the right context to feed into models right now. And currently, that's a lot of copying and pasting." - Product Manager

"I've been desperately trying to figure out how to best automatically manage my AI's context (primarily from Google Docs, Jira, Figma for design files and my product scrapbook, and Granola for notes) without having to explicitly upload PDFs or call out specific file names proactively. I have not been able to crack it though." - Product Manager

“The approach of ‘use AI as your assistant with broad knowledge so it can do everything’ just hasn’t worked. I end up spending too much time trying to give it enough context to think like I would.” - Product manager

These people are incredibly AI-forward and hardworking. Some work hard to do this manually: dedicated time blocks, scrapbooking, and lots of copy and pasting. These people’s time is rare and expensive, so that tells us how valuable that is.

Normal people can’t afford to do that. Imagine how many could benefit if we made this slightly easier.

We want everyone to leverage their hard work—the thinking and writing scattered across their tools—with AI, no extra effort.

Who might want this?

The answer today: AI-forward geeks like us who use AI locally (Claude Cowork, Cursor, Claude Code, Codex, OpenClaw etc).

Long term, we’ll make this less geeky (local AI agents are just our starting point). We'll also bring it to teams and enterprises.

“I want peace of mind that my own self and my company are not operating with amnesia, and that they're constantly getting better at the fastest rate they possibly can.” - Founder

Why is this free, open source, and offline?

We envision a world where every knowledge worker runs Familiar (or something like it). Getting there requires being free, open source, and offline.

We're inspired by Peter Steinberger, whose approach to OpenClaw showed us that “having nothing to lose” is a great strategy. We’re building this because no responsible company in their right mind (e.g. with revenue and legal teams) would build this. (Maxim and I frequently answer each other with “WWPSD.”)

We believe in positive-sum games. A world where every knowledge worker runs Familiar (or something like it) is a world with abundant opportunities.

How can I get involved?

We’re onboarding 30 alpha testers over the coming week

It's up on GitHub, though we'd love 10 minutes to onboard you.

Book 10 min with Maxim

Not ready to book? Hit reply. Tell us you’re interested, tell us we’re nuts, tell us you want Windows, whatever’s on your mind. I read and respond to everyone.

Join our open-source community

The repository is here. Nothing makes nerds like us happier than a ⭐️ on Github.

I can't imagine working without AI sitting over my shoulder, watching everything

I know that’s a weird sentence, but AI needs wayyyyyy more context, and that context is so much bigger than my past chats or MCP integrations.

I'm getting answers that would be impossible even for the highest-paid coach. No human can watch my screen all day, remember everything, and provide piercing answers on demand. For any amount of money.

With Familiar, AI can.

We currently don’t have another cohort scheduled. We both have things cooking that might affect when we’re able to do this again.

Aman texted this yesterday- I can’t give details, but I can say he’s putting his money where his mouth is.

My top 54 PM AI guides, in one place, for free

In the year since my podcast with Lenny I’ve published 70+ AI tutorials for product people, all for free (I love making these). Since there’s so many, I’ve organized my favorites by career goal, below.

I’m feeling overwhelmed just looking at this, so remember 1) most are a few minutes and 2) follow your curiosity and click on what brings you joy.

I don’t paywall knowledge, so all the links below are yours. Tomorrow’s workshop offers outcomes.

For some of you, the knowledge is the outcome: you’ll work through this and learn by trial and error. For those who want structure, efficiency, and support our workshop provides:

• Butt-in-seat accountability and our trained team to get you unblocked, to ensure you get through all of it

• The freshest stuff, (I learn faster than I can publish, and you also get Aman who I learn so much from) plus curated external resources and a Slack community for ongoing discussion.

• 10 live hours with Aman Khan and me, abundant open Q&A, and my personal Whatsapp forever (yes me, no not 🦞)

Get $375 off tomorrow’s cohort with promo code ”outcomes”. You can also book an AI build sprint for your team.

Whether you join us live or dig into the links, I’d love to know know how to make this more valuable for you - hit reply for my personal gmail.

-Tal

PS If you’re currently unemployed or a university student, Aman and I have something for you.

Build your AI product sense

• 🔥 How to build AI product sense - Our guest post in Lenny’s Newsletter that’s… not a post. You’ll see.

• The eval that saved our butts - Behind the scenes of building and testing our Lenny’s Newsletter interactive AI experience

• How does ClawdBot/OpenClaw work? - A short video course breaking down how 🦞 works, specifically for non-technical product people.

• What do AI agents “see” when they use a browser?

• I wanted to understand subagents, so I had Claude Code write standup comedy and workshop it with three AI audience members

• I wanted to know how “compaction” works, so I did brain surgery on Claude Code - What happens when AI agents run out of context window

• I noticed that Claude Code and Cursor have started managing their own context - What I saw when I had AI stockpile my transcripts before the impending Granola paywall

• Dissecting Bolt’s system prompt - Peek behind the magic to see how AI prototyping tools actually work through careful prompt engineering (with Aman Khan)

• How I built ChatGPT, extended thinking, and web search from scratch

Stay current on AI (without the screaming)

• I don’t aspire to be an AI engineer, but I do want to understand one layer down - here’s who I’m following

• The firehose of AI announcements - how to filter AI announcements and decide what actually matters

• Defy the AI-hype-industrial complex: introducing the People’s Post Generator

• How I’d explain AI agents to an 8-year-old

• ”Agents” are a Whatsapp group chat where the smartest member has the memory of a hamster

• I’m frustrated with the jargon around AI agents

Build your AI thinking partner and personal OS

• 🔥 Build your personal AI copilot - Canonical step-by-step guide - free guest post in Lenny’s Newsletter

• Why build an AI thinking partner - Understanding the value of building a personal AI thinking partner for product management

• Onboard your AI thinking partner - How to set up and onboard your AI thinking partner with the right context and instructions

• Give AI your user flow - Give your AI thinking partner visual context about how your product works

• Gossip to your AI thinking partner - Keep your AI thinking partner updated with every conversation and project update using the gossip workflow

• Prompting is the fastest way of transferring knowledge as a product leader

• My most-used phrase with AI is “ask me one question at a time”

Move away from chat UIs and into AI coding agents

• Intro to Cursor with a simple use case - How PMs can use Cursor when ChatGPT and Claude are blocked at work

• AI thinking partner architecture in Cursor - Three levels of context for building your PM thinking partner

• Why Cursor hits different than ChatGPT projects - Small UX changes that create big workflow impacts

• Don’t use Cursor to start - Why you should begin with Claude or ChatGPT projects first

• What Cursor taught me about building AI products - Why the right interaction design matters more than the model

• PM frameworks as Cursor rules - Encode your favorite PM frameworks into reusable, systematic tools (with Tristan Rodman)

• Why Cursor instead of Claude Code? - Why choose Cursor over Claude Code, and when not to (with Patrik Bostrom, Zev Arnovitz, Sven-Erik Nielsen)

• Context & hallucination management - How to manage context effectively and prevent AI from making things up (with Hilary Gridley, Sven-Erik Nielsen, Tristan Rodman, Zev Arnovitz)

Use AI agents for real work

Strategic work

• Ask your codebase dumb questions - How non-technical PMs can understand technical complexity (with Hilary Gridley)

• Synthetic users - Using AI to generate synthetic user personas for product testing

• Am I prioritizing correctly? - Using AI to challenge and pressure-test your prioritization

• Mining support tickets with MCP - Query support tickets conversationally without pre-built dashboards (with Sven-Erik Nielsen)

Productivity

• Vibe code your Personal OS - 5,952 product people signed up for our free lightning lesson.

• Claude Code is creating my tasks for me

• Turn meetings into action items with Cursor rules - Automatically extract action items from one-on-ones and add them to your todo list (with Patrik Bostrom)

Prototyping

• AI prototyping prompts - Prompts for turning product thinking into working prototypes

• From PRD to prototype in minutes - Transform finished PRDs into interactive prototypes that look like your product (with Zev Arnovitz)

Data analysis

• Data visualization with AI - LLMs excel at rapid visualization prototyping - generate and iterate on HTML/CSS visualizations in seconds (with Hilary Gridley)

• MCP for product data analysis - Using Model Context Protocol to connect AI to databases (with Dan Byler)

• Reverse engineer a competitor’s revenue - How to estimate competitor revenue from accidentally leaked information using rigorous cohort analysis (with Hilary Gridley)

• Recommend strategic metrics - Define the right activation metric and move it using app store reviews and qualitative data (with Hilary Gridley)

• Analyze behavioral overlap between features - Find overlap between features to build smarter user journeys and increase AI feature adoption (with Hilary Gridley)

• How LLMs analyze data - Understanding how LLMs use code to perform reliable data analysis and combine quantitative with qualitative insights (with Hilary Gridley)

• When to trust AI product data analysis - Best practices for balancing AI-driven analysis with product instinct and expert validation (with Hilary Gridley)

Soft skills and hard conversations

• Why inner game matters more than ever - A strong emotional inner game is a force multiplier for leveraging AI at work (with Jack Cohen)

• Separating yourself from your emotions - A practical technique for gaining clarity before difficult conversations and decisions (with Jack Cohen)

• The Chihuahua Method - A framework for managing overwhelming emotions at work (with Jack Cohen)

• Inner dialogue before a meeting with AI - Use AI to simulate and understand difficult emotions before challenging conversations (with Jack Cohen)

• Use AI to simulate a hard conversation - Use AI to rehearse difficult stakeholder conversations - practice being both yourself AND seeing through their eyes

• Practice hard conversations with AI voice mode - Practice difficult feedback conversations with AI playing a defensive recipient (with Jack Cohen)

• Rehearse for Big Tech interviews using AI - Use AI to prepare for high-stakes tech interviews and difficult conversations (with Ben Erez)

• 🔥 Product manager is an unfair role. So work unfairly. - Guest post in Lenny’s Newsletter

Phew! I’m thankful to love what I do, and that it resonates with all of you.

• Tomorrow Hilary Gridley, Aman Khan and I are banding together to bring you a free Lightning Lesson to give you a repeatable system for validating AI ideas cheaply: How to Know What AI Products to Build.

• On Thursday, Aman and I are embark on our second cohort of Build AI Product Sense.

  • Thanks to Lenny choosing us for “Top course in product,” you get $375 off and $1,395 in free credits.

  • Thanks to Aman being….Aman, we’re giving away mac minis.

In 21 days Granola paywalls my AI agent’s access to my own transcripts. Watching Cursor & Claude Code export them showed me how AI agents proactively manage their own context.

Exporting all my meeting transcripts is a deceptively difficult challenge for an AI agent, because MCP responses go into the context window no matter what the next step is.

This is fine for “what’s the weather” but for an entire meeting transcript that just needs to be saved to a file, that’s really inefficient and unnecessary. Since I don’t need the LLM to engage with the transcript (just save it to a file), this quickly fills up the context window for nothing.

Ideally you’d be able to chain MCP tool calls to skip the context window, so the output of “get transcript” gets piped straight into “write file.” (This is how Clawdbot/OpenClaw works btw.)

These inefficiencies ended up being a gift because I found myself with a context-intensive task that would overwhelm Opus 4.6’s context window. Kind of an AI “will it blend.” With an impossible task at hand, I asked both Cursor and Claude Code to export all my Granola transcripts locally, same prompt, both using Opus 4.6.

Both surprised me! While they took different approaches, I loved that both made “context management” part of their plan.

💠 Cursor’s approach

For both tools, I first enabled the Granola MCP. Because the paywall is still three weeks away, I still have access to the “get transcript” tool call.

Cursor proceeded to:

• List transcripts in Granola’s server

• Fetch & save each transcript one at a time

• Context window runs out quickly

• Auto-compact the context to keep going

Until this point, it’s what I expected. But I noticed that “clear space in my context” was part of the agent’s plan. It was doing it strategically! (reminds me a lot of Claude playing settlers of Catan for 75 minutes)

Notice the line “…once I clear space.” How freaking interesting is that?

👾 Claude Code’s approach

Claude Code proceeded to:

• List transcripts in Granola’s server

• Dispatched 6 subagents in parallel, each fetching 25 transcripts

--- that’s it ---

Ok fine there was one last step… it took forever (yes, with 6 agents in parallel, more below).

It was so cool to see both products take on the same task, same model, different tactics. The differences are not as interesting as what they have in common: both agent harnesses are super aware of their own capabilities and formulating strategies around their own context window.

They know what their own limits are, and they know what their tactics and strategies to handle those limits, just like they use any other tool.

If you insist I look at the differences, I don’t necessarily think that one strategy is better than the other. Think about it: “compacting your chat thread over and over again” isn’t very different from “spawning a subagent.” The only difference might be that you can run subagents in parallel, which didn’t matter here since the bottleneck was Granola’s rate limits. So, maybe Cursor knew what it was doing.

(Another note about the differences: this is not a systematic evaluation. If I ran this 10 times, they might each employ both strategies equally.)

So much to think about in this seemingly routine task! Go do this while you can - more for the learnings than the loot.

P.S. Credit to for doing this first and tipping me to the fact that we’ve all been put on free business trials (so the MCP server is still, for now, giving us access to the “get_meeting_transcript” tool).

• On Monday, , and I are banding together to bring you a free Lightning Lesson to give you a repeatable system for validating AI ideas cheaply: How to Know What AI Products to Build.

• On Thursday, Aman and I are embark on our second cohort of Build AI Product Sense. Thanks to Lenny, you get $375 off and $1,395 in free credits.

Looking forward to seeing you at one or both!

Two days ago, and I published our guest post in Lenny’s Newsletter that is deceptively not a post, but instead one big prompt for Cursor to teach you Cursor from inside Cursor. (If you haven’t experienced it yet, do that now, then come back.)

Before sending it out to 1.3 million subscribers, we ran usability studies with about a dozen people. We sent them the instructions and asked them to record themselves going through the process. We then watched each recording end to end (we were not gonna let AI get in between us and this gold) and made incremental improvements between each round.

The feedback leveled off and we felt good. By the time Lenny gave it a spin, we thought we’d seen everything.

Famous last PM words…

Of course Lenny discovered a new bug. Of course it was the worst one of all: the LLM found itself in a weird existential crisis, the “game over” kind with no clear way to recover, because it’s misleading the reader as well.

The lazy PM voice in my brain went, “Meh! one in a dozen… how often would this actually happen?” The less-lazy PM voice in my brain said, “You could simulate this a bunch of times automatically and see.”

In a normal product, Aman and I could use an observability platform to copy Lenny’s chat thread and re-run it at scale. Unfortunately, since it’s a open-source experience that runs locally on each person’s laptop, we gave up all analytics and traces.

We did however, have Lenny’s screenshare video. I manually reconstructed Lenny’s AI chat thread by repeatedly hitting pause and copying the text on his screen.

Here’s how I got a sense of how often it was really happening:

1. I recreated Lenny’s conversation inside an eval platform (any of them work). The left side is the simulation: what Lenny pasted in, what the LLM said back, Lenny’s first response, the tool call, the faked tool response. I rebuilt the whole thread up to the moment right before it went haywire.

2. I created 10 duplicate rows to start, which would allow me to run 10 alternate universes where Lenny hits that point.

It was happening a lot.

Ok bug, you’ve caused this PM to get up from his chair and…. edit the prompt! Minus getting up from my chair.

Now, I could just look at the results myself and see if the changes worked… or I could create an automated eval to do it for me.

1. I created an “LLM as a judge” or “Model-based grader” that would evaluate the result. As the subject matter expert, I had the eval assert “is it having an existential crisis” (not in those words, see video).

2. I kept tweaking the eval prompt and seeing if it evaluated correctly until I trusted it.

3. Now that I trusted my eval, I could now iterate on the post’s prompt itself. I didn’t start with any professional-grade prompting, more just clarified and removed wishy-washy language.

4. Each time I made a change, I re-ran the 10 alternate universes. When those were done, I ran the eval on each of the 10 results.

5. As the original prompt got better and better, the results started to look more and more green. After a bunch of iterations, I got to 9/10 passing, which is way better than what I had before.

6. From here, I could get more certainty by raising the number of parallel Lenny universes.

All this to say, I’d love to hear of any weird experiences you’re having (email me at talsraviv at gmail.com). Maybe there’s another eval Aman and I need to build! If so, like Lenny, we’ll probably ask for your help reconstructing your trace.

-Tal

The post you’re about to read took over 100 hours to create. That’s because it’s not a post. It’s an open-source interactive AI experience that will help you build AI product sense.

Tal and Aman ran dozens of usability sessions, wrote evals, optimized each prompt you’ll find below, and even partnered with Cursor to get you free credits (see below!) so that you can try this at home.

I’ve never seen anything like what they’ve put together, and I’m excited to bring it to you.

This was ridiculously fun to create. Aman and I PM’d this like an AI product, and not one of those hours felt like work.

The flaw: we have no observability. Since this is an open source product that runs locally on your laptop, we have no traces, and no analytics. The only user feedback we have is YOU telling us what sucks. Just reply directly to this email to reach me, and I’ll cc Aman. We’re hungry to know!

You can stop reading here, and go experience it for yourself.

If you’re still reading, I’ll share that after our first cohort hit #1 on Maven across all categories, last week Build AI Product Sense was selected by Lenny for top course in product, which means you can use the code LENNYSLIST for 15% discount and $1,395 in free credits for Superhuman, Linear, Sprig, Dovetail, and Gamma.

Ok enough, go enjoy the post. Excited to hear from you. -Tal

That sentence feels exciting and terrifying at the same time.

I’m going to skip the superlatives. Instead, I’ll make four contradicting statements:

1. There’s no magic ingredient in OpenClaw that didn’t exist before.

2. It’s like nothing you’ve ever experienced.

3. It’s incredibly dangerous and risky.

4. It absolutely should inspire a generation of companies.

All of those things can be true.

You might have also seen the “AGI!” screaming on social media. No one knows what “AGI” means, but if you define it as being able to be on-the-go and texting an AI employee back at the office who can can do anything a human can on a computer, that’s what this is.

The best part is, it had nothing to do with “insane new AI model” or even agent harness. It was just a matter of breaking unspoken rules, and giving unreasonable permissions that responsible people decided not to do until now.

Meaning, all this was possible, but no legitimate company would dare release something like this. It took an open source project by a financially independent entrepreneur (Pete Steinberger) to say “fuck it” and lob this thing into a crowd of unrestrained software developers.

🛑 Please don’t implement anything in this post until you read the section about security, at the end.

These videos are for non-technical product folks

I made this as a product person who wants to understand one layer down. Our job as product people will be to make this technology (or some form of it) accessible to a lot more people than it is today. To do that, it’s really important to build a strong AI product sense for how this product works.

This post is a free mini-course in OpenClaw for non-technical product people. We’re going to take this apart together, step by step, so you have a deep intuition for how this works and how to apply it to your customers.

Quick note: I filmed these videos as “ClawdBot” but Anthropic’s lawyers made the (free, open source) project change its name mid-recording. It’s now called “OpenClaw.”

Special thank you to Maxim Vovshin, , and for their contagious, hands-on, hacker energy.

Understanding how it works

First, I let Cursor hold my hand and be my technical tutor throughout

OpenClaw is difficult and esoteric to set up. The docs are a mess. It’s a huge barrier to entry for somebody non-technical. I later realized this is good friction given how dangerous this is. This should not be easy.

So, I probably shouldn’t say this, but I got around the complexity by letting my AI coding agent handle everything for me. I could even ask dumb questions throughout the process.

To do this, I let Cursor clone the open-source repository and run it, rather than running it according to the website instructions. This ended up saving my ass, see “security” section at the end.

I asked Cursor to explain things to me slowly

Since it’s an open-source project on GitHub, I could give the URL to Cursor and have it copy the code to my computer. Anytime I needed something explained, I could ask Cursor to read the code and spell it out for me slowly.

What I did here:

• Cloned the GitHub repository into Cursor so I could explore the code

• Before running anything, indulged my curiosity by asking Cursor to explain the technical elements to me (gateway, agent loop, WhatsApp connection)

• Learned it’s just three existing things combined: Claude Code + more permissions + optional Whatsapp/Telegram

• Understood the system prompt structure (identity, tools, skills, memory)

Ingredient #1: The Claude Opus 4.5 model

It’s funny to write “Claude Opus 4.5 isn’t new” because it was released 63 days ago. It’s a much better model in many ways than its predecessors, particularly on the being-good-at-calling-tools-and-reasoning-its-way-around front.

Even though it wasn’t 10X better, it was a big step forward and it enabled a lot of things that were not possible before, in an exponential way.

Ingredient #2: Agent harness (in this case, Pi)

This is the program on your computer that keeps calling the LLM API. it holds the chats, it calls the tools, and it keeps calling the LLM any time something new comes up so the LLM can give its response and decide what to do.

This is the same as your local coding agent. For example, Claude Code is available to developers as an SDK. All the cool features you see in Claude Code (skills, subagents, basic tools) are available to anybody building their product.

Pi, by Mario Zechner, is a super minimalist version of this.

Ingredient #3: Unreasonable, unsafe levels of access

Claude Code can only run commands in your terminal—and that’s already risky. OpenClaw is that plus even more abilities, like being able to run your browser with all your authenticated cookies. You can even give it access to your password manager 😱.

We’ll talk about the full ecosystem of skills and tools a bit later.

Ingredient #4: Whatsapp/Telegram/Slack with it from your mobile phone, on the go

I originally assumed this was a gimmick. I eventually came around to truly appreciate how boss it feels to step away and just text with the thing.

How does the AI agent have Whatsapp/Telegram? if you’ve ever used Whatsapp desktop or Telegram desktop, you’ve had to go through the process of scanning a QR code with your mobile app to authorize the desktop app. There’s an open-source library that allows any desktop app to show you a similar QR code, and receive the same permissions as Whatsapp/Telegram desktop apps have.

You have two options:

1. Give it its own phone number (that only you know)

2. Have it use your Whatsapp/Telegram account

While the latter is easier, remember you’re granting an autonomous AI agent access to all your DMs and group chats, and trusting it to only read the ones you told it to.

How the ingredients mix together

There’s a long-running process on your physical laptop that listens to Whatsapp messages. When those messages arrive, it sends them to Claude Code (just like you’d hit enter in a chat box). Claude Code can then execute a ton of tools it was never meant to have.

You can read the system prompt

At the top of every agent thread, there’s a system prompt. AI models tend to weight that more heavily just as a result of their training. It’s just text in a context window, but it’s the first thing there.

OpenClaw’s system prompt is a big template with lots of blanks to fill in:

• Identity (who it is)

• Tools available (what it can do)

• Skills (capabilities it can use)

• Memory (what it remembers about you)

• Current date and time (so it knows when it is)

• Permission to give silent replies if it has nothing to say

Understanding the building blocks means you can think about recombining them yourself. It’s good models + simple agent loops + lots of tools.

You can build with these same ingredients for your customers.

Setting up OpenClaw

What I did here:

• Let Cursor run the installation commands for me

• Used Cursor to explain confusing onboarding options (hooks, skills, memory)

• Set up personal WhatsApp (it has access to all my messages, trusts me to only send relevant ones)

• Added allow list so only I can trigger it

• Skipped scary options I didn’t understand yet

Until now we were just letting Cursor read the code, and hadn’t run anything. Let’s set it up!

I put Cursor in agent mode and told it to set it up for me. Again, Cursor is my technical tutor holding my hand here. Cursor can run all the necessary commands and it can see the output if they succeed or fail and decide to try something else. (All of the docs are also part of the open source repository, so Cursor could read those directly.)

Cursor asked me if I wanted it to take care of the onboarding or I wanted to go through it myself. I decided to experience it directly. Still, I ran the onboarding in a terminal inside Cursor, so I could still keep the left side of my screen to be a Cursor agent chat. That way, I could ask all my dumb questions if a technical term in the onboarding didn’t make sense, or I didn’t know how to answer a question

For example, I had no idea what “hooks” were in this context. So I asked Cursor. Turns out hooks are nice things like:

• Session memory

• Startup tasks that happen when you boot it up

• Heartbeats: tasks that run on a schedule (we’ll get to this later)

It asked for my OpenAI API key so I can use Whisper. And my 11 Labs API key so it can talk back to me in voice. I was happy to fetch both.

This onboarding was not very self-explanatory. But using an AI coding agent as my tutor made it accessible. I basically asked “what the hell is this?” at every step.

First signs of life

With onboarding done, it sprung to life and loaded up a web chat (I hadn’t set up texting yet).

What I learned here:

• Bootstrap.md is the “birth” file (used once, then deleted)

• Memory is just text files (identity.md, user.md, soul.md, memory files)

OpenClaw said, “I’ve got bootstrap.md here.” So I turned to Cursor and asked, “What the hell is a bootstrap.md?”

Cursor searched through all the files and explained that when OpenClaw starts up, if it sees the existence of this file, it knows it’s the first time it’s been run. Bootstrap.md is used once as instructions to interview me, save those answers in other files, and then deleted.

My answers are stored in long-term memory as, you guessed it, text files! We’ve got an agents.md, a soul.md, an identity.md, a user.md. They could all be in one file, but they’re separate because it’s nicer (?).

Fine FINE it’s magical

No matter how much I tried to deconstruct OpenClaw into its ingredients, it was impossible to stay skeptical when it started doing things.

What I did here:

• Requested a flight search

• Watched it control my browser and iterate

• Saw how it converts HTML to an accessibility tree to understand the page

“Show me what you can do,” I said to OpenClaw. It came back with a list of abilities. Browser control caught my eye.

Fine. “Can you search for a flight for me on Google Flights from Tel Aviv to Los Angeles in early March, after March 7th.” And…. it launched Chrome… and it did just that.

It took control of my computer. It was trying things. If you’ve ever watched an AI coding agent at work—iterating, trying something, get the result back, try again—it felt the same, but now it’s broken out of the terminal.

Waking up on its own: “heartbeats”

What I did here:

• Set up a flight price tracker that runs on a schedule

• Learned heartbeats are cron jobs for AI

• Understood it only works if your computer is on (hence Mac Minis)

• Watched it edit its own heartbeat.md file with scheduling logic

I saw the list it gave me also had “heartbeats,” so I asked it to show me using my flight search example.

Heartbeats are scheduled processes. Once in a while, it’s going to check Google Flights. It edited the heartbeat.md file. It decided to keep some memory here of what the prices were. It came up with alert logic.

There’s a reason very few frontier LLM agents run autonomously without a human hitting enter: it can get very expensive, very fast. “Heartbeats” lets you do it anyways.

Your computer needs to be always on

I asked: does my computer need to be on for this to work?

Yes. This is why people are purchasing computers to run this (real or virtual). I initially rolled my eyes watching and discuss their plans to buy Mac Minis and then…

My mind is running wild. But I’m trying to stay grounded, reminding myself: this is Opus, which we’ve had, wrapped in Claude Agent SDK, which is just Claude Code, wrapped in OpenClaw, which just gives access to anything all the time.

This is getting super expensive, super insecure. And I love it.

OpenClaw’s skills

What I learned here:

• Anything you’d use your computer for can become a skill: 1Password, Apple Notes, Notion, GitHub…. it can even run Claude Code (that’s meta) or Codex etc.

• A worldwide community of developers is hard at work building skills for you

Since OpenClaw wraps the Claude Agent SDK (which, as you remember, powers Claude Code), it is built to use Skills! (A Skill is a saved prompt that an LLM can decide to pull in.)

There’s also a whole community of people developing these:

Every time I see another software app available as a skill, I feel excited and horrified at the same time.

Basically anything that you would use your computer for, you can install these skills. It’s gonna take me a while to wrap my mind around the geometric explosion of use cases this allows.

Texting with OpenClaw on Telegram

What I did here:

• Set up Telegram as another channel (with Cursor’s help)

• Asked it to search for flights from my phone via Telegram

• Watched it control my computer while I was away from it

Telegram is a little bit easier to set up than Whatsapp. I asked Cursor for help and it walked me through it step-by-step.

And then… well, putting that in words feels silly, the video captures this better.

It is amazing how much the same ingredients that have existed can be rearranged in a way, with a little bit of not giving a fuck and nothing to lose, can just do magical things.

Voice messages

What I tested here:

• Sent a voice message

• Watched it transcribe with OpenAI’s Whisper model

• Received a voice message back (ElevenLabs text-to-speech)

• Recognized this is more existing ingredients combined

Um, I sent it a voice message, and it responded with a voice message.

Part of me wants to let go and enjoy the magic. Another part of me just keeps deconstructing it and saying: all these things existed. Voice messages, 11 Labs text-to-speech, Whisper speech-to-text, Claude Code, Opus. They’re just being all put together.

Goddamnit, it’s magic.

Security (boo! I mean yay!)

What I did here:

• Ran the security audit with Cursor (THIS IS NOT ENOUGH)

• Learned about prompt injection risks (THIS IS THE LEAST OF OUR WORRIES)

• Decided to treat this as inspiration, and then shut it off (TURNS OUT IT DIDN’T REALLY TURN OFF)

I told Cursor I was worried about security. What could make this more secure? What are the risks that I should be aware of?

Cursor told me there’s a security audit I should run regularly, and so I did. That’s not enough at all.

I’ll bottom line it: there’s a ton of risks you can see and don’t see. It’s easy to make a tiny mistake and leave yourself (and your employer) very exposed.

Oh, and a day later it turns out I didn’t really turn it off. I found out by accident, only because it launched my “flight search heartbeat” I happened to schedule.

I was sooooooo glad I was working with OpenClaw with Cursor as my intermediary, I have no idea what I’d do otherwise.

If there haven’t been any large-scale disasters attributed to AI until now, OpenClaw is 100% where they’re going to come from. But also, we can’t not try. Hashtag humanity.

The right way to do this could be a separate computer with very limited permissions. Maybe give it a separate user. Maybe give it access to some accounts.

Or just don’t use it. Sometimes the right answer is “this is cool, let’s learn from it, now let’s turn it off.” (That’s gonna take a lot of self-restraint.)

What to do with this

OpenClaw is a beautiful source of inspiration that should make a lot of product teams ask “how can we bring this magic to our industry in a way that’s way more accessible than this and way more secure?”

All the ingredients were there in the open. It just took somebody who had nothing to gain and nothing to lose to say “fuck it” and combine all the ingredients without worrying about the consequences.

Now the world has one big, fat proof of concept done as an open source project in public, that should not be used by most people for most of the time.

Install it. Understand it. Get inspired. Turn it off.

Then go bring this magic to your customers in a way that’s actually responsible.

If you found this helpful, you definitely know that AI may have made building cheap, but knowing what to build has never been harder. , , and I banded together to host a free Lightning Lesson to give you a repeatable system for validating AI ideas cheaply, so you can distinguish between novelty and lasting value.

TLDR: When Cursor decides to open a webpage with its built-in browser, the browser tool converts the page into a clean, bulleted outline before passing it to the LLM.

How I saw it for myself

1. Make sure you’re on the latest version of Cursor (they recently updated this feature).

   

2. Enable the browser tool in Cursor by visiting Settings > Tools and turn on “Browser tab”

   

3. Make sure your chat is in “Agent” mode (and choose Claude Opus for its tool-calling prowess)

   

4. Ask your LLM to open a webpage in a browser tab:

   Please open https://www.worldhappiness.report/ in a cursor browser tab (no other way)

5. If that doesn’t work, cajole it.

6. If that doesn’t work, firmly insist.

7. &@%#$

8. Look at the “browser tool call” in the chat, click into it, and open the file.

9. You’ll see the “webpage snapshot” file - that’s what gets passed to the LLM context

What the LLM “sees”

In that file, you’ll see “YAML” format, which is a (relatively) human-readable outline of content. This preserves HTML’s hierarchy while stripping away the noisy tags, javascript, and CSS that aren’t relevant to the LLM.

(YAML stands for “YAML Ain’t Markup Language.” …nerds).

To be clear - the LLM isn’t distilling the webpage. The browser tool is doing the work. An LLM would be a slow, expensive, and imprecise way to convert HTML to YAML. This is a perfect task for old-fashioned, deterministic code.

To summarize, the browser tool converts the webpage to YAML and puts it in the LLM’s context window. That’s it. Once I saw that, I could stop thinking of “AI browsing” as something magical and mysterious.

However, YAML isn’t the only way that agents look at webpages. "Claude for Chrome" alternates between text extraction and taking screenshots. I’ve seen both Cursor and Replit take screenshots while vibe coding to verify visual, frontend changes.

Which method is better? It’s a tradeoff. From Anthropic:

Browser use agents require a balance between token efficiency and latency. DOM-based interactions execute quickly but consume many tokens, while screenshot-based interactions are slower but more token-efficient.

For example, when asking Claude to summarize Wikipedia, it is more efficient to extract the text from the DOM. When finding a new laptop case on Amazon, it is more efficient to take screenshots (as extracting the entire DOM is token intensive).

(“DOM” means “webpage HTML” in this context)

You can empower the agent make the call. That way, it might try the YAML route, see thin results, and decide that screenshots are a better choice:

In our Claude for Chrome product, we developed evals to check that the agent was selecting the right tool for each context. This enabled us to complete browser based tasks faster and more accurately.

I could read Anthropic’s blog posts all day, but Cursor lets me see it for myself in my daily use. It's an agent harness running right in front of me!

If you’re on a product team, or work with product teams, set aside 3 minutes to play with this. It’ll build your AI product sense better than any newsletter.

-Tal

Subagents are the equivalent of you opening additional side chats while using ChatGPT.

Say you’re in a ChatGPT conversation (e.g. plan a surf trip), but you need to research something on the side (e.g. what wetsuit am I ok with). You don’t want to derail the main thread, so you open a new tab, ask your question there, get an answer, then bring just the bottom line back to your original conversation.

That’s a subagent. When Anthropic built “subagents” into Claude Code, they automated that plumbing. (If this sounds a lot like calling a tool, you're right. Manus calls this "agent-as-a-tool")

“Fresh side chats” are valuable for (at least) three reasons:

1. Keep side quests from polluting the main agent context window (e.g. you just need the answer, not the research/discussion that got there)

2. Get a fresh pair of eyes. If you want code review, or an eval, or a second opinion, you don’t want the same context window that came up with the answer to also judge it. (Zevi Arnovitz did this manually long before it was a feature.)

3. Protective layer against prompt injection: When accessing a risky source (random website), the main agent can do it via a subagent. If the source turns out to be malicious, the subagent has a lot fewer permissions so the downside is low, and insulates the main agent.

My plan was to have Claude Code spawn subagents, then observe it closely. Spoiler alert: it’s all text files and context windows.

To give Claude Code a reason to spawn subagents, I created a comedy writing skill (a “skill” is a saved prompt) that tells Claude Code to write a joke, then spawn three audience members to react to it.

While Claude Code ran, I watched from two camera angles:

1. The terminal, where I could see it spawn a “writer” subagent, then three “audience” subagents

2. The actual files on my computer where Claude Code stores its memory.

That second part is what demystified it for me. I navigated to the hidden ~/.claude/projects folder and watched Claude Code create:

• A main thread (the comedy club?)

• A comedy writer subagent

• Three audience subagents

• Another comedy writer subagent to act on their feedback.

Even though these look like scary “JSONL” files, they’re just text files containing chat history, that are human-readable if you open them with this prettifier extension.

Watching the files appear in real time made it click for me. Before this, “subagents” sounded like something sophisticated I was missing. Now I see it’s just chat threads spawning chat threads.

I recommend taking apart Claude Code while we still can. It helps to see that subagents aren’t magic: they’re just opening another chat thread, getting an answer, and bringing just the bottom line back into the main thread.

In other words, text files all the way down.

TLDR my rant: Most AI content induces fear and FOMO. Influencers and algorithms symbiotically profit from keeping us impressed and confused. calls it a “race to the bottom.” We’re in an AI-hype-industrial complex.

The solution is for real operators to share how they’re really using AI at work. Together, we can drown out the "this model is INSANE" posts, magic-spell prompts, hasty infographics, and "comment to get my pdf/prompts/belly button lint."

If you feel inhibited: It doesn’t have to be novel or fancy. We're all hungry to know what you're doing, what's working—or not—for you.

Except operators don’t have time. Not to write posts, and definitely not to “learn social media.” I asked you to post more, with no way to do it.

So I’m freely sharing the AI Skill I use to write my social posts. “The People’s Post Generator” is an AI Skill, basically a folder of text files that an LLM can pull into conversation when relevant.

The People’s Post Generator

I’ve been growing a little folder of text files to help me write my social posts. It helps me take a messy idea (Slack screenshot, Whatsapp chat export, Granola transcript, or just speech-to-text rant) and sculpt it into a post I’m proud of.

I feel proud of posts that are personal and practical. My process goes like this:

1. Get excited about something I experienced, and make a screen recording (e.g. Loom, Screen Studio, Cap, etc.)

2. Copy the transcript into AI chat (Cursor in my case) and tag the @peoples-post-generator folder

3. AI then asks me one question at a time to get more context out of me, deciding together on the focus, and drafting the post.

4. I share back my drafts, and AI holds me accountable to my own bar for publishing.

These text files grew through use, not Prompt Engineering™️. Each time AI flubbed, I added a line of instruction. If I came across an example I liked, I’d copy-paste it in.

Slowly, it became useful to me. When friends mentioned social posts, I'd zip the folder and send it over. After the seventh time, I decided to share it with the world.

Quick start guide

A Skill is a folder with text files:

You can read each text file yourself. It contains:

• A lightweight process

• Bank of examples (non-cringe “hooks,” posts, calls to action)

• Guides to good writing (how not to sound like AI, a directive to write like you talk, and a TLDR of The Elements of Style)

You can use it in with favorite LLM:

1. Download the zip file to your computer

Head over to the People’s Post Generator Github page and download the zip file:

2. Give it to your LLM of choice

ChatGPT/Gemini tutorial

These steps are the same for Gemini, using “Gems.”

Claude web tutorial

Cursor tutorial

These steps work for any VSCode-based AI code editor (e.g. Antigravity, Cline, etc.)

Claude Code/Claude Cowork tutorial

3. Give AI a messy idea

I love ’s commandment to “document, don’t create.” This AI skill is perfect for raw records of conversations and experiences. Throw any of these into the chat box and ask it to use the skill to help you turn it into a post:

• Screen recording transcript (my favorite)

• Slack thread screenshot

• Whatsapp chat export

• Meeting transcript ( did this mid-call)

• Rant into your favorite speech-to-text app

• Or a nearly-finished post you wrote on your own

AI will figure out where you are in the process and continue the conversation from there.

4. Make it your own

This skill works out-of-the-box. However, the folder make-it-yours/ is well, completely yours. Once it’s on your computer, feel free to delete anything, add your own ideas, or remove examples that you don’t like. (In fact, you can edit the whole thing, SKILL.MD, included.)

5. [Optional] Contribute improvements

Most of the changes you’ll make to this will be personal (examples you like, your personal bar for posting something). However, you might find ideas that are relevant to other people using this skill (whatever their taste). For example, a better conversation experience, or a way of making this skill more likely to get picked up by AI.

If you find a universal improvement, I would love for you to open an issue or even make a pull request:

1. Open a coding agent (Cursor, Claude Code, etc.)

2. Tell it:

I want to make a pull request to this github repository but not sure what to do: https://github.com/talsraviv/peoples-post-generator

Let AI hold your hand through the process.

Power to the humans

I created this for myself, gradually and organically. I needed something that would hold me accountable to my own standards when I'm tired and tempted to just hit publish.

Now that it belongs to everyone, real people can call me out when I don’t follow my own rules, not just AI.

If AI skills are new to you, the best way to learn is by doing: try the tool. And if you want to learn how to evaluate this kind of fluency in PM interviews (or demonstrate it yourself), I'm co-hosting a free Lightning Lesson with Ben Erez and Aman Khan:

* Please don’t infer my political views from this post.**

** Please don’t infer my political views from this footnote.***

*** I’m staunchly pro-footnotes.

I realized that chat LLMs aren’t very conversational. (Think about the last time ChatGPT answered with a wall of text.)

When I collaborate with an LLM and it answers with a novel, it reminds me of that junior person who gets a task and runs off for two weeks in a cave, working super hard on it. They come back and do a velvet reveal.... that’s completely misaligned.

The longer the LLM’s response, the more assumptions it’s making. For example, imagine if an LLM replies with 8 questions for you to answer. If your answer to the first question isn’t what it expected, the other 7 become irrelevant.

So, when I say “ask me one question at a time” I’m giving AI a chance to align and recalibrate after each question.

This also extracts more context out of me that I wouldn’t have articulated upfront, and gives me tons of chances to apply my taste along the way.

Whether I’m writing a spec, drafting a research plan, analyzing data, working on positioning, or just drafting an update: I get way better results, way faster, when I make it a conversation.

It’s the same thing we already know about collaborating with humans: frequent alignment beats cranking away in isolation.

It did this by frequently and intelligently pruning its own context. Watching this made me realize that a big part of “long-running agents” will be a new term called “context editing” (also known as “compaction” also known as “$%@&-I-ran-out-of-room-please-summarize-this-thread-for-me-so-I-can-start-a-new-one”)

A few days later, I noticed that Claude Code’s /compact command allowed me to add optional custom compacting instructions!

This new (?) feature reminded me a lot of the Catan demo, which felt like an important learning opportunity as a PM. I don’t aspire to be an AI engineer, but I do want to understand one layer down.

I realized this option in Claude Code allowed me to be a partner in the “context editing” process: I can help decide what to summarize vs. preserve granularity. When Claude Code has a new feature, I pay attention, because features that start in Claude Code tend to show up in what Anthropic offers to developers.

I figured it’s important as a PM to see for myself how this works inside Claude Code. But to do that, I’d have to crack it open.

I had no clue how to do that, until and gave me the key hint: Claude Code stores all your conversation history as text files on your computer.

Below are the steps I took to watch compaction happening with my own eyes. If you’re a product builder looking to build your AI product sense, I highly recommend making time to try this out yourself.

If it feels intimidating, remember: it’s all just text files on your hard drive.

Here’s how to experience this yourself:

1. Create a new empty, throwaway directory anywhere on your computer

2. Open it in Claude Code and give it a Sisyphean task (I told it to Create a separate text file for each word of the Gettysburg Address, pinging me in between every file.)

3. While that runs, visit your user’s root directory (one directory above your Desktop, aka ~)

1. Press Command+Shift+Period (Mac) to show hidden folders. In Windows file explorer click View > Show > Hidden items

2. Click on .claude/ folder

3. Find the projects folder and click on the one that corresponds to your throwaway directory, above (This is the behind-the-scenes of Claude Code’s brain for something you’re working on, not the actual folder you’re working on)

1. Use Cursor/VSCode (free) to open that folder.

2. Install the VSCode JSONL Gazelle plugin (open source, thank you !)

3. Open the most recent JSONL file you see there - each row is a message in your conversation!

4. Run the /compact command in Claude Code with any custom instructions, e.g. Preserve the last 5 turns of our conversation.

5. Watch what happens in the last rows of that file as you compact.

What I learned: When you compact, Claude Code doesn’t just summarize and discard your conversation. It keeps your conversation, and just draws a line - creates a “compact boundary” - in the conversation history. Then, it writes a summary of what happened before.

This way it keeps the full original conversation. The new thread can still retrieve any details from before compaction if needed. That is so damn cool.

Why this matters: What we get in Claude Code is similar to what Anthropic provides to developers building AI agents - so inspecting our tools is a great way to build product intuition for these technologies.

The best way to understand LLM applications is to reproduce them inside coding agents (like Cursor or Claude Code), and then look inside. It’s all text files.

-Tal

Most people use ChatGPT like a chatty Wikipedia or a Slack update generator... and wonder if they’re missing some secret prompt that only influencers know.

There’s a different way: treating AI as a thinking partner, a teammate, a colleague who knows your company strategy, remembers that conversation with your manager from last Tuesday, and can challenge your assumptions when everyone else has gone home.

When you get the foundations right (context, mindset, daily habits) AI becomes a force multiplier for everything you do.

Once you use AI daily for real work, you develop X-ray vision for AI products. Where others see magic, you see architecture.

I’ve helped over 20,000 product leaders build their intuition for AI applications this way, including at Apple, Google, Amazon, Microsoft, Meta, Netflix, and Stanford.

Enrolled students include:

• Director of Product at Google Labs

• Head of Engineering at AWS

• VP Product at Fidelity Investments

• CPO of Okta

• CEO of Zapier

My students tell me they use their AI thinking partner every day, and it’s always open at work. They move so fast that leadership taps them to teach everyone else how they got so productive. (Two weeks after one student set this up and joined a new company, his CEO remarked, “It feels like you’ve been here for months.”)

Most importantly, they become that person in the room who actually understands AI applications. They build a reputation for saying ‘let me show you’ and demoing live. When Anthropic or OpenAI announce something new, they intuit it was coming because they already solved that exact problem last week.

How students use AI as a thinking partner:

• Analyzing messy qualitative insights (while staying close to the raw data)

• Simulating a hard conversation with a stakeholder (and then reverse-simulating it)

• Uploading a meeting transcript and getting coaching + turning it into a simulation

• Querying and navigating databases thoughtfully

• Understanding their team’s codebase (even if non-technical)

• Asking “what’s the single most important thing I should do right now?”

• Getting challenged on their thinking when everyone else has gone home

• Spinning up functional prototypes with plain English

• Building tailored AI workflows for their team’s actual pain points

• Making context engineering a team sport so when one person levels up, everyone does

You don’t need magic spell prompts or social media hacks. You just need a quiet moment to get hands-on.

1. Follow my free tutorial to build your AI thinking partner

Follow my free, hands-on tutorial in Lenny’s Newsletter for getting started in ChatGPT projects, Gemini Gems, MS Copilot, or Claude projects:

Lenny's Newsletter

Build your personal AI copilot

👋 Welcome to a ✨ free edition ✨ of my weekly newsletter. Each week I tackle reader questions about building product, driving growth, and accelerating your career. For more: Lenny’s Podcast | How I AI | Lennybot | Lenny’s Reads | Courses | Swag…

Read more

9 months ago · 669 likes · 16 comments · Tal Raviv

2. Paste this prompt into Cursor to build your AI product sense

Now that you understand the fundamentals, follow my and ’s free hands-on interactive tutorial to move to Cursor and build your AI product sense.

Lenny wrote in his intro: “…it’s not a post. It’s an open-source interactive AI experience that will help you build AI product sense… I’ve never seen anything like what they’ve put together, and I’m excited to bring it to you.”

Lenny's Newsletter

How to build AI product sense

👋 Hey there, I’m Lenny. Each week, I answer reader questions about building product, driving growth, and accelerating your career. For more: Lenny’s Podcast | How I AI | Lennybot | Lenny’s Reads | Favorite AI and PM courses | Favorite public speaking course…

Read more

2 months ago · 357 likes · 32 comments · Tal Raviv and Aman Khan

PS If you’re interested in the “making of”, check out: the eval that saved our butts

3. Go deeper with dozens of free AI video tutorials I’ve published on this Substack

I also have guest posts in Lenny’s Newsletter about building AI automations and one of the most popular Lenny posts of all time: Product manager is an unfair role. So work unfairly.

4. Stay in the loop

Get notified when I publish new tutorials and guides:

Subscribe now

5. Go deeper with an AI build sprint

If you got this far, you’re in the top 5% of people leaning into AI agents for real work (even if it doesn’t feel that way, trust me).

To go deeper, join me for a live Maven cohort. To get your team on board, let’s plan a build sprint.

About me

Hi, I’m Tal 👋 I was Patreon’s first growth PM, where my team increased paid conversion by 31% over tens of millions of consumers, virally reached 9 million social feeds in one week, and our methods became part of the Reforge curriculum. I was also the founding PM of the Patreon API, which powers products at Spotify, Reddit, and Discord.

At AppsFlyer, I PM’d the engine powering TikTok’s mobile links, and the data feeds powering Amazon. Most recently, I was Riverside’s first PM, leading an AI product used by Spotify, Marvel, TED, NYTimes, and Apple to interview guests like Mark Zuckerberg, Michelle Obama, Bill Gates, and Taylor Swift.

I started my career by co-founding a profitable SaaS company, and working with the founders of Wix and DuckDuckGo.

But here’s what I left off that resume: 18 months after ChatGPT launched, I still wasn’t using AI at work... like at all.

My team and I were building AI products used by tens of thousands of people. But when it came to using AI in my own job, I was a proud Luddite.

When I tried using ChatGPT, I found it disappointing for strategic work, more like consulting a chatty Wikipedia. Deep down, I felt discouraged that I didn’t know the exact magic spells of prompting that only influencers seemed to know.

Then I hit a wall at work. That’s when I finally figured out what was missing: context, mindset, and daily habits that elevated AI to a teammate.

Since then I’ve been a guest on Lenny’s podcast and featured 3X in the newsletter. My workshop with Aman Khan, “Build AI Product Sense,” hit #1 on Maven, with students from Stripe, Figma, Nike, JP Morgan Chase, McKinsey, LEGO, and every company in MAANG.

These days I try to share as much as I can for free, and you’ll find it all here. If you want to go deeper or bring your your whole team along, let’s plan an AI build sprint.

Outside of work, I volunteer as a surf instructor for people with disabilities 🏄‍♀️

Reach me on LinkedIn or talsraviv at gmail.com.

(For me, one layer down means “building applications on top of foundation models” to borrow from Chip Huyen’s awesome O’Reilly book, AI Engineering)

This is an incomplete list, highly biased to my own fumbling around the internet, and whose emails I’m excited to open. I’d love your recommendations on technical people I should be following!

🏢 Company engineering blogs

• Anthropic’s blog is full of accessible, widely-cited bangers. Subscribe in the footer here. Even if you don’t, engineers will forward these to you.

• Whenever I wonder how Claude’s consumer features work, I can usually find it thoroughly explained in their developer docs.

• Surge’s blog is thought-provoking real talk, backed by their hands-on experience. (I can’t figure out how to subscribe, since they don’t have an RSS feed. In the meantime, I keep refreshing)

• Also following HumanLayer, PromptLayer, and Amp’s company blogs.

🤓 Individual engineers

• Doug Turnbull for agentic search (plus a bajillion lightning lesson recordings)

• Eleanor Berger and Isaac Plath put out incredible hands-on walkthroughs

• I like how Ben Guo (Zo Computer) thinks.

• Armin Ronacher early at Sentry, creator of Flask is refreshingly honest and accessible

• Mario Zechner exploring agents in public

• Jesse Vincent is open-source-famous and created the Claude Code Superpowers plugin. Anytime i see a new post I know it’s going to be accessible and super hands on.

• Simon Willison is one of the best at working in public and focuses on making everything practical.

Many of these don’t have email forms. I use a free RSS-to-email service to email me when they post new stuff (I chose Blogtrottr after doing zero research).

😵‍💫 How I discover good stuff

I stay away from feeds. Luckily, there’s people (and AI) who browse social media so I don’t have to. For technical stuff:

• I love the concept of Hacker News digest newsletters. Instead of compulsively refreshing HN, these free services wait a bit, see what shakes out, and email you a summary: https://hackernewsletter.com or https://www.hndigest.com

• Ben Tossell ends each newsletter with a section called “What I’m reading,” a solid way to discover high signal essays.

I also make time to go down rabbit holes. I schedule reading for mornings when my brain is freshest (esp earlier in the week). That’s when I’m naturally curious and excited to click on additional links. 

👓 Happy reading!

You don’t have to be an engineer to follow engineers. When Anthropic/OpenAI/Langchain/etc. announce a new capability (for developers or consumers), I’ve already seen these folks thinking about it in public.

It feels good to anticipate instead of chase.

What I felt as I've been using Cursor is the ingredients are the same, but they're arranged slightly differently. There's slightly different UX because it was built for software development. These small changes add up to very big phenomena in how I work.

If anybody's experienced El Niño, you know that some very small changes in the speed of the wind in a particular part of the ocean changes a little bit, and then the whole globe is completely different.

One small change with Cursor: the project knowledge is editable. The enormous impact is that it gets really smart, really fast. When I was using Claude projects, I had my project knowledge and lots of chats, then I had outputs, and then I would decide which of those becomes project knowledge. In a Cursor environment, the project knowledge just keeps updating in a very tight loop.

Another small change: the initiative doesn't live in a thread like it would in a project. It lives in a doc that just keeps getting better. I've noticed that my chat threads in Cursor become super disposable, super throwaway. In an LLM project, the history matters. But in Cursor, the history is stored in a document. These chats are just little tools that you use to do something and throw them out. So you never really hit a limit.

Another really cool thing: you can use initiatives as context when you're talking about another initiative if you want. They can know about each other. Even the hierarchy in which you organize folders is context to Cursor. That also tells it how things relate to each other.

**➡️ Cursor accidentally built an excellent PM thinking partner. Your knowledge evolves in real-time, initiatives cross-pollinate, and chat becomes disposable because everything important lives in your documents.**

Zevi Arnovitz shows how to transform finished PRDs into interactive prototypes that look enough like your actual product to effectively communicate with your team.

"We've talked a lot about using Cursor as a thought partner. Now I want to show how we can go from thinking and creating docs to creating prototypes. It's easy to translate what we're thinking to our team," he explains.

Zev's setup includes his copilot with general editor context, people on the team, competitor research, and initiatives. "We noticed that users were having really a hard time selecting palettes for their site. We gave all kinds of default palettes, but users wanted their own colors, their brand colors. So we decided to create a product that allows them to upload a logo or an image and generate some palettes from that."

After finishing his PRD, he tells Cursor: "Based on this PRD, I want to create a prototype. I want to put it in the demo hub we've created together. I want the flow to be where users can open from the side panel of the editor the design panel. They can upload their logo, and then we'll be able to generate 3 different palettes for them."

He adds photos of both the editor and how palettes are shown within the editor for context. "It's a bit long, but that's pretty much it. I added the PRD for context. Now I'm going to let this run."

His demo hub showcases all his prototypes in a clickable format: "It's a cool way to evangelize this and show this to the team." The resulting prototype looks similar enough to the Wix editor that users can upload a logo (like Starbucks), generate palettes, select them, and see the content on stage change.

"This is way overkill, a lot of stuff that Cursor decided to do. But this is just a one shot prompt. It took me maybe 5 minutes. You could definitely do a few iterations and make it really look like what you want. It really simplifies translating this to the team."

After writing your PRD, give it to Cursor with screenshots of your actual product. Ask for an interactive prototype. Five minutes later, you have something clickable to show your team instead of static mockups.

Tristan Rodman shows how to encode PM frameworks into reusable Cursor rules for systematic interview synthesis and PRD creation.

"The frameworks folder has a number of Cursor rules based on activities that are outlined in some of my favorite PM lit," he explains. "You can think of these rules as specific instructions for an LLM to go about a task. Same systematic approach you'd use when applying these frameworks for your product work. It's a prompt you can save and reuse."

Following Continuous Discovery Habits by Teresa Torres, he created an interview guide and mocked up user interviews. "It's really hard to replace the intuition and practice of doing these interviews, but synthesizing snapshots after the fact, that's where I need a lot of help. If I don't do that in the meeting it never gets done."

Using his "create interview snapshot" rule, he processes three transcripts. "Using the at key to bring in different documents really feels like you're just tagging in different bits of context." The rule produces methodical interview snapshots for each participant.

"You can start to see some trends are emerging. Everyone's frustrated by time signatures. Cool, makes a lot of sense." He chains documents through, using another rule for synthesizing the interview snapshots into a full readout. "I brought the whole folder in. We're gonna produce a whole research readout."

From there, he creates a PRD with another rule. "Make a PRD for adding time signatures. One of the parts of the PRD rule is it asking me for some clarifying questions. Some questions about scope, some questions about visualization UI design. For the purposes of this demo, I'm being super lazy, and I've told Claude, use your best judgment based on the interviews."

➡️ Create Cursor rules for your favorite PM frameworks. Tag interview transcripts with @, run your synthesis rule, chain the outputs into research readouts, then generate PRDs. Your frameworks become reusable, systematic tools.

Sven-Erik Nielsen shows how to use GitHub Copilot and MCP servers to interrogate support tickets for themes and insights without pre-built dashboards.

"As PMs we all know that support tickets are just an incredible source of insight, but sometimes it can be a little bit difficult to get them or the interfaces may not be conducive to really fluid analysis," he explains.

Sven uses MCP (Model Context Protocol) servers to connect to databases. "I already installed MCP servers beforehand. There's lots of good literature out there on the Internet how to do that." He starts them up and puts his query in the chat pane using agent mode, which is key for having the agent manipulate files and create changes to markdown.

His prompt: "I want you to reason over this storm events table and help me understand the themes and insights. I don't want you to do the work for me of determining what I should know and think about these incidents. Think of these as though they are support tickets. I need you to be kind of a thought partner here."

The LLM connects to the database through the MCP server and goes through a series of steps. "It's kinda cool because you can look at basically the chain of thought. That's a good way to double check you're getting the right information, and you can use that to iterate and improve your prompt."

The output reveals themes: infrastructure vulnerability, tree damage, homes that are vulnerable, water hazards, meteorological patterns. "It gives me a sense of these incidents. It does give me some specific instances to go back in, like these event IDs here, to drill back into the support database."

"That way I don't have to be beholden to a pre-built dashboard, which kind of constrains my thinking. On the other hand, sometimes, while I love writing SQL queries, it can be not the most fluid way of trying to think through what's going on in the data. So it's kind of a hybrid of those two approaches."

➡️ Install MCP servers to connect your IDE to databases. Query support tickets conversationally instead of writing SQL. Get themes and specific ticket IDs to investigate without being constrained by pre-built dashboards.