If you've been here a couple weeks, you'll remember "the room" from Issue 6 where we walked through what the AI can remember during a chat, and how to keep it from getting cluttered.
This issue is about what to put in that room. Most of us spend our AI hours tweaking what we ask. "Maybe if I rephrase this. Maybe if I say it nicer. Maybe a longer prompt would help." The AI gives us something. We tweak. We tweak again. And another twenty minutes gone.
But Anthropic put it this way: using AI isn't about finding the right words and phrases for your prompts. It's about answering a bigger question: what context setup will get the model to behave the way you want?
Translation: stop polishing your prompt. Build up your files.
How you ask matters, but less than you think. Wording isn't nothing. A 2024 study did show that changing only the format of a prompt (the same words with a different layout) can significantly swing performance on some tasks. But that effect shrinks as AI gets bigger and better. With today's models, wording matters less than people realize.
What you give it to read matters more than you think. The thing that turns a mediocre AI answer into a useful one is hardly ever the way you write your prompt. It's what the AI knows when you ask. When you start a new chat, the AI doesn't know your voice, your project, your last six decisions, your client's preferences, what you tried last week, or what you ruled out. You do. That gap is the whole game.
You wouldn't ask a contractor to design your kitchen renovation when they've never seen your house, never met your family, and don't know you can't stand fluorescent lighting. That's how most of us are using AI.
Here are three files you can put in front of the AI this week.
A. A few examples of your own writing. Not a description of how you write, but examples. Three pieces you'd want the AI to sound like: a blog post, a long email, a draft you're proud of. It's like sending the AI a portfolio instead of describing your portfolio. Drop them into one document with a heading at the top of each one. Use it whenever you need the AI to draft in your voice. The AI learns more from one example than from twenty adjectives.
B. Everything about one project you're working on. The brief, the current draft, the decisions you've reversed, what's still open. The real stuff. But don't summarize. When you open a chat about that project, the AI walks in already knowing what you're inside.
C. Recent decisions, with the why behind them. Two or three sentences each. I picked X because Y. I stopped doing Z because W. This is the file most people are missing, and the one that could make the biggest difference. The AI can see what you've kept. It can't see what you've ruled out. Without this file, it'll keep suggesting things you settled weeks ago. The AI will reference it like notes from a meeting it was in.
Here's a snippet of what File C might look like:
# Decision Log
## Cut the long version of every closer (5/1/26)
The closer was running 80+ words. Trimmed to 30. The last line
lands harder when there's less around it.
## Stopped writing openers I don't have (5/1/26)
Tried opening Issue 7 with a personal scene that wasn't completely true. I didn't connect with it, so I cut it. Started with what I noticed instead.
## Started keeping this file (4/29/26)
Two issues in a row I'd forgotten a constraint I'd already settled. Logging decisions and the why so future-me has the receipts.You don't need all three today. Just start with one.
Whichever you build, save it as a markdown file. I keep mine in a folder on my desktop, but your notes app, a Google Doc, or a Dropbox folder works too. The point is that you don't rebuild it next time. If you don't have a place you already keep things, Claude and ChatGPT both have a feature called Projects that pre-loads files into every new chat.
More isn't better
Volume isn't the point. Choosing what's relevant is. It's like packing for a trip: a bigger suitcase doesn't mean you should bring everything you own.
Chroma's 2026 research tested 18 different AI models, and found every single one got worse as the pile grew. Anthropic's own guidance, from the same engineering post, points the same way: give the AI the smallest set of stuff that gets the job done.
Remember: How you ask matters less than you think. What you give the AI to read matters more. More isn't the same as better.
The test for any document you're about to add to a chat: "Is this specific to what I'm asking right now, or am I dumping it in because I have it?" If it's the second one, leave it out. The room still gets cluttered, even when it's bigger.
me+machine