Why I stopped writing prompts and started writing files

For the first few months of using Claude properly, I lived inside the chat box. I’d open a new conversation, paste in a long prompt I’d written the week before, and start working. The prompts lived in a Notes file, sorted by date, prefixed with whatever project I was on. By month three I had about forty of them.

It wasn’t working. Three small things kept going wrong.

First, I kept losing my prompts. Notes is a search-and-find tool, not a working memory. If I wrote a good summarising prompt for client reports in February, by April I’d forgotten the exact wording, and I’d write a slightly worse one instead. Two months apart, same task, measurably worse output. The work was getting better, slowly, but the institutional memory of how it got better lived in my head, not in the system.

Second, every conversation started from zero. Claude didn’t know who I was on Monday, even though it had been talking to me on Friday. I’d waste the first five minutes of every session re-establishing context. The system felt smart but never seemed to learn anything.

Third, the prompts were too long. The good ones were eight hundred words. I was copy-pasting eight hundred words to start a conversation about a forty-word task. That’s a smell.

The fix, when I found it, was embarrassingly simple. Stop writing prompts. Write files.

The pattern is called CLAUDE.md, and it’s the most underused feature in the Claude ecosystem.

How it works

Anywhere Claude is operating — a Project, a folder, a workspace — you put a file called CLAUDE.md, in plain markdown. Claude reads it automatically when it starts. It contains the standing instructions that would otherwise live in your prompt: who you are, how you write, which tools you use, which files matter, what good output looks like.

You write that file once. From then on, every conversation starts pre-loaded with context. The eight hundred words don’t disappear, they just live somewhere stable. Notes goes quiet.

What mine looks like, in skeleton

At the top level of my Claude workspace, my CLAUDE.md is short. It says I’m a regulated payments operator in Melbourne, that I write in Australian English with contractions and short sentences, that I prefer Claude to ask one clarifying question before producing a long output, and that for any task involving numbers, I want the working shown. That’s it. Ten lines.

Inside a project folder, the CLAUDE.md gets more specific. For the Pesabase work, the file lists the three reconciliation tools we use, the standard format of the partner statements, and the one situation (settlement breaks above five thousand dollars) where Claude should stop and flag for human review. Twenty lines.

Inside a specific task folder, it’s just the task. “Draft the monthly board memo. Use the template at /templates/board.md. Length: one page. Tone: brisk.” Five lines.

When I open a chat about the board memo, Claude has all three files loaded in order. It knows I’m a payments operator, that this is the Pesabase project, and that I want a one-page brisk memo from the template. I’ve typed nothing.

What changed when I switched

The day I switched, the work changed. It didn’t get more impressive. It got more boring, which is the actual sign that an AI workflow is working. Outputs stopped being little fireworks of “wow Claude can do that” and became reliable, predictable, slightly improved versions of what I would have produced myself. Claude stopped being a party trick and became infrastructure.

There’s one thing I want to be honest about. The CLAUDE.md hierarchy isn’t a magic upgrade. It’s an investment. Writing the top-level file took me a couple of evenings, and I rewrote it three times before it settled. Writing project-level files takes a slow afternoon per project. The skill is mostly in figuring out what to leave out, what context Claude actually needs, versus what I’m including out of nervousness.

But the payoff is real. The system compounds. Every new project starts with the cumulative learning of every previous one. That’s the thing prompts in Notes can’t do.

If you try one change this week

Open Claude, find the Projects feature, pick the project you use most, and add a CLAUDE.md file with ten lines about who you are and how you work. Don’t try to write it well. Don’t try to anticipate everything. Just put down ten true sentences. Use the system for a week. Then come back and edit the file.

In a month you won’t open Notes again.