My AI agents read my notes more than I do
I think out loud with ChatGPT, but Claude Code and Codex do the work, and they can't see my chats. One note is the handover. Here's the five-minute habit.
TL;DR
- I do my thinking out loud with ChatGPT. My actual work gets done by Claude Code and Codex, and they can’t see my ChatGPT chats.
- So when a conversation changes what I’ll write or build, it becomes a note in my notes folder, where every agent can read it.
- My agents now start most jobs from notes like that. They do about 90% of the reading and nearly all the writing. My job is checking their judgement.
My best thinking was stuck in a chat
On a Sunday morning in August, I spent a long drive talking to ChatGPT. I was trying to put into words why I keep all my notes as plain text files rather than in an app like Notion, something I’d been circling for months.
By the end of the drive, it had finally clicked. I had a clearer view of what I wanted to write and what I wanted to build.
It also lived in one long ChatGPT thread, tangled up with every wrong turn I’d taken to get there.
That’s a problem for me, because ChatGPT isn’t where my work happens. My work happens in Claude Code and Codex: AI agents that read the files on my computer, do the job and write the results back. They can’t open my ChatGPT history. As far as they were concerned, that drive never happened.
One note is the handover
So the conversation became a note in my notes folder. The note is longer than it needs to be, but the top of it does the job:
Notion may be the better app. For the kind of second brain I want, Markdown is the better substrate.
My knowledge base is core infrastructure for how I think, work and learn, and increasingly how I use AI. So I want the source of truth to be independent of any one app, maintainable as it grows, and safe for AI agents to work on.
Below that sit the arguments, the objections I expect, and the questions I still haven’t answered.
About a week later, my agents used that note to draft a long essay. Three weeks later, a different agent session used it to plan a workshop. I didn’t have to explain the idea again either time, and neither agent ever saw the ChatGPT conversation.
You’re reading the short version of that essay.
Why my agents read more than I do
My agents do about 90% of the reading in my notes, and nearly all of the writing. That’s a rough estimate, but the direction is obvious. They start each job by reading the relevant notes, decisions and instructions, and they finish it by updating them.
I read the parts that matter: the decisions, the claims, and anything a client or a product will rely on. I fix a note when it has missed the point. The agents do the filing, and I do the judgement.
In June I wrote that documentation has a new reader. My notes have the same new reader, and it reads a lot more than I do.
That’s also why the notes live in plain text files I own, rather than inside one app. Any agent can read them. I can see every change an agent makes, compare it with the last version and undo it.
Try it with one conversation
Pick one AI conversation that changed what you plan to do. Ask an agent to turn it into a short note that answers three questions:
- What did I work out?
- Why does it matter?
- What do I want to use again?
Ask it to link back to the original chat, in case the detail matters later. Read the note, fix anything it got wrong, and save it somewhere your agents can reach.
Next time you start related work, point your agent at that note first. It starts from your best thinking instead of a blank page.
Start with the conversation you’d be most annoyed to lose.