Rowan Cheung, founder of The Rundown AI, recently shared an interesting way to use the enormous pile of conversations we’ve been accumulating with AI assistants: export the whole thing and ask the model to audit you. Not summarize your chats. Analyze the patterns across them.
What do you keep asking about?
Which decisions do you repeatedly revisit without making?
What do you claim is important versus where your attention actually goes?
What aren’t you delegating, and why?
I tried it. This was slightly unfortunate.
Apparently, I Have a Problem
The analysis didn’t find any shortage of ideas. Quite the opposite.
My problem appears when I need to choose one project, put it in front of actual humans and discover whether anybody wants it.
Then I suddenly become extremely sophisticated. Let’s compare another option. Let’s calculate the risks. Maybe improve the prototype. Perhaps reconsider positioning. One more feature wouldn’t hurt. Maybe this other project has better economics.
You know. Research.
Except sometimes “research” is a wonderfully respectable word for: I’m slightly terrified that I’ll launch this thing and everyone will discover it’s crap. Annoyingly plausible.
Rowan apparently got his own collection of unpleasant observations: problems delegating, a tendency toward control, and a habit of burying himself in work rather than dealing with uncomfortable conversations.
So the method appears democratically insulting.
Try It Yourself
First, export your complete conversation history from whichever AI assistant you’ve been using.
With ChatGPT, use the data export option in Settings. Claude has an equivalent export under Privacy. Other assistants generally have some version of the same thing.
Once the archive arrives, extract the conversation history and turn it into something your model can ingest.
My archive was around half a gigabyte, which is a slightly disturbing amount of evidence against myself.
Then start a fresh conversation and give the model your history with this prompt:
You have all of my conversations. Don’t summarize them. Compare them against each other.
Find contradictions between what I say is important to me and what I actually spend my attention on.
Identify:
questions I repeatedly ask in different forms
decisions I keep circling around without making
tasks I avoid delegating
what I may be trying to protect by refusing to delegate them
At the end, give me three well-supported conclusions about myself that I am most likely to strongly disagree with.
Base them on recurring patterns, not isolated examples.
No flattery.
You can obviously modify this depending on what you’re interested in.
Then Keep Digging
The first audit is only the interesting part because it gives you hypotheses.
After that, ask things like:
What weaknesses keep appearing across my conversations, and what should I do about them?
What do I repeatedly ask you to do that should become a reusable workflow instead?
Which parts of my work barely use AI today even though they obviously could?
Which decisions consume disproportionate amounts of my attention?
Where does my stated strategy contradict my actual behavior?
That last one can be particularly entertaining.
Why I Think This Works
Most self-analysis has a terrible dataset.
You.
More specifically, your memory of yourself. And we’re remarkably talented at rewriting that dataset.
Your AI history is different.
If you’ve been using these systems heavily for several years, you’ve accidentally created a longitudinal record of: problems, ideas, decisions, anxieties, unfinished projects, recurring questions, things you repeatedly postpone, and whatever particular nonsense was consuming your attention on a random Tuesday fourteen months ago.
The individual conversations aren’t particularly important. The recurrence is. And LLMs happen to be quite good at finding patterns across messy text. So give your chatbot the accumulated evidence and let it finally complain about you.
You’ve been feeding it your problems for years.
It deserves a turn.