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

Because chat LLMs aren’t very conversational

Wispr (my speech to text app) says my most-used phrase with AI is “ask me one question at a time.” It got me to reflect on why I say this so often.

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.

Opus responded with a single spot-on question, then paused for my response.

Without that phrase, Opus printed four mostly irrelevant options (each with pros and cons), plus a section it titled “my instinct” and ended with 3 questions for me, all in the same breath.

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.