The first time I used Claude, I tried to build myself an executive coach. The kind that actually knew me. My values, my voice, the stuff I never said out loud.
I spent hours meticulously feeding it everything about me: work docs, project plans, old journals, financial spreadsheets, years of files dragged into one chat.
Then I asked Claude about a partnership I'd ended last year… and it didn't even know the partnership existed.
How could it not know? I had given it all of my journals!
Turns out Claude had been quietly compressing what I was feeding it the whole time, without warning or asking. It was summarizing the documents into paragraphs, or even sentences. Whole files reduced to post-it notes on a desk.
For a long time, I tried to fix this the obvious ways. Better prompts. "Please remember this." Longer system messages. Begging. Waiting for the next model to be smarter.
None of it worked. I'd give Claude all the context, things would go well, then partway through a long chat it would start drifting: forgetting my tone, suggesting things I'd ruled out, getting more generic.
I assumed I was prompting wrong.
Turns out the mental model was. I was treating my AI like a brilliant collaborator with a memory problem. The issue had nothing to do with memory. It had to do with where my AI works.
The AI lives in a room. Everything in the room is what it sees. Everything outside doesn't exist.
But the room isn't the AI's whole brain. The AI walks into the room already brilliant. It's read most of the internet, knows history, code, and how a mortgage works. It knows the world. It doesn’t know you. The room is for the you stuff.
Three things to know:
Every word inside the room is real to the AI. Every doc you paste, every previous turn, every line you wrote 20 messages ago. It all shapes what comes next.
Open a new chat and the room is empty. Yesterday's project, last week's prompts: gone, unless you bring them back.
And the one nobody told me—the room gets cluttered fast. Things at the edges get summarized into post-it notes and the things in the middle blur. The AI never tells you. It just keeps working, slightly worse, from whatever's left on the desk.
This is what happened to me on day one. The system summarized to make space, and Claude kept working off a blurred version without telling me what was lost.
If you've heard the term context window, that's the room.
Once I stopped trying to fix the AI and started managing the room, everything changed.
Two moves did most of the work:
Move 1: Bring a brief into the chats that matter
Five to ten lines. Who you are, what you're working on, the voice you want, what to avoid. Save it. Paste it into any chat that matters.
Here's mine:
I'm a writer. I publish a newsletter called me+machine about how everyday people can actually use AI in their lives. Practical, warm, never preachy.
My voice is conversational, specific, slightly self-deprecating. I lead with stories, not theory. I trust the reader. I do not write in corporate-speak. I do not use "leverage" or "unlock" as verbs. I do not use exclamation points unless something is genuinely surprising.
When I ask for help on a draft, I want clear structure, specific examples over adjectives, and a willingness to push back when an idea is mid. If a section isn't working, say so.
If I haven't told you the project, ask before assuming.
It's a little bit like 50 First Dates. Drew Barrymore wakes up every morning not knowing Adam Sandler. He doesn't get frustrated. He doesn't try harder. Instead, he builds her a video — every important thing she needs to know to step back into their life — and they pick up where they left off.
Your brief is that video. You're not catching the AI up because it's broken. You're catching it up because that's the loving, normal way to work with someone who lives one chat at a time.
Move 2: Save the room before it gets compressed
This is the move almost nobody uses. It would have saved my executive coach experiment on day one.
Long chats get worse over time, walls or no walls. Answers get vaguer, the AI drifts from your original instructions, and things you said up front quietly lose weight.
When you do hit the actual wall, the system summarizes for you. Claude does this automatically when the context hits around 95% capacity. It's called compaction, and it’s happening whether you notice or not. But you don't have to hit a wall for the room to start losing focus.
Don't let the system decide what survives. Beat it to the punch.
If you've discussed something important, ask your AI to write it all down. Current state of the project. Decisions you've made. Your tone notes. Open questions. Have it produce an MD file or a clean recap. Save it somewhere you control.
A few weeks ago I was deep in a long draft when I noticed Claude getting fuzzier. Suggesting things I'd ruled out. Losing the voice. So I asked Claude to dump everything important into a recap doc, opened a fresh chat, and pasted it in.
The next response was sharper than anything I'd gotten in the cluttered one. Rooms get rearranged. Snapshots don't.
Stop hoping the AI will remember. Start deciding what's in the room with it.
Welcome to room management.
This is Issue #6 of me+machine. Every week: one way to make AI work harder for you. Nothing you need a tech background for.
Your turn. What's something you wish your AI knew about you that it doesn't? Hit reply.