Skip to content
Get in touch
← Notes

Getting AI rightOctober 8, 2026 · 3 min

Ask your chat history what you keep getting wrong

Export your chat history and ask it what you keep getting wrong.

By Chris Battis

THE SAME ASK, AGAIN

A friend of mine writes grant proposals for nonprofits, freelance, eight months into using ChatGPT for a first draft on every one. She likes her process. She also rewrites the opening paragraph on almost every proposal, and had never once asked why.

Last week she exported her entire chat history (Claude and ChatGPT both let you do this from Settings) and dropped the file into a new chat with one question: what do I keep asking for help with, and where do I usually give up.

The answer took thirty seconds. Forty-one conversations, out of maybe ninety, included some version of "make the opening less generic." Forty-one times she'd hit the same wall and fixed it by hand instead of fixing the prompt.

She'd never have found that by remembering. She found it by asking the one source that had actually been keeping score.

You don't remember your own patterns. Your chat history does

Nobody tracks their own AI usage the way they'd track anything else that mattered to their business. You wouldn't run a business for eight months without looking at the numbers, but most people run ChatGPT for eight months without ever looking at what they actually asked it, what worked, and what they gave up on.

The pattern is invisible in the moment because each instance feels small. One generic opening, fixed by hand, forgotten by lunch. It only becomes visible in aggregate, and nobody has the aggregate sitting in their head. The model does, because you told it, conversation by conversation, for months.

It's the same gap you see in a single prompt: the AI doesn't know your business unless you tell it. Turn the question around and the same gap exists at the level of your own habits. You don't fully know your patterns unless something shows them back to you.

The audit. Fifteen minutes, most of it waiting on a download

Step 1. Export your data. In Claude: your name in the bottom left, Settings, Privacy, Export Data. In ChatGPT: the icon bottom left, Settings, Data Controls, Export. Both take a day or so to email you the link. Request it now, do the rest tomorrow.

Step 2. Open the download, find conversations.json, and drop it into a new chat with one of these:

What tasks do I ask for help with repeatedly, and how could I
get a better result from those tasks?

Find the conversations where I got frustrated, gave up, or ended
things before I got what I wanted. What went wrong, and how
should I have prompted differently?

What are the ten most important patterns you can see in how I
actually use you?

Step 3. Read what comes back before you argue with it. The instinct is to explain away the pattern. Don't. If it names something you do forty-one times, that's not an outlier. That's the actual shape of your workflow.

What you do with the answer is the part that compounds

A pattern by itself is trivia. The move is to fix the prompt that keeps producing it, once, so the fix applies to every future proposal instead of the next one.

My friend added one line to the top of every grant-writing chat: open with the specific outcome this grant produces for a named beneficiary, never with our organization's history. One sentence. Forty-one rewrites, gone from the next batch.

Run the audit again in two or three months. New patterns show up as your work changes; old ones you already fixed drop off the list. It's not a one-time diagnostic. It's closer to a checkup: worth repeating on a schedule, not worth obsessing over weekly.

Same audit works whether you live in Claude, ChatGPT, or both. Export each separately and run the same three prompts against each file.

Together, we'll uncover what AI can improve.

It starts with one conversation about your business.

Get in touch
The AI advisory of Chris Battis
Ask a question
Portsmouth, New Hampshire