My AI workspace: from 31 KB of instructions to 3 KB
When CLAUDE.md becomes a novel
In March I explained how CLAUDE.md changed the way I work. That is still true. But in seven months, the file at the root of my projects had grown to 31 KB: the status of every app, the history of every session, dozens of rules.
The result: at every session, the AI read all of it before starting and drowned what mattered in everything else.
A router instead of an encyclopedia
I reorganized everything into a small system I call ARMS:
- CLAUDE.md becomes a 3 KB router: who I am, how I work, and where to find the rest
- An index per folder (projects, content, business, personal) says where everything lives
- App sheets and session history live in their own files, read only when needed
- An inbox receives deliverables to sort
Nothing was deleted: everything was moved as is, with a backup of the original.
The harness matters as much as the model
Two days later I came across an idea that clicked: the "harness", everything around the model, matters as much as the model itself. And you improve it in small, tested layers, not by adding paragraphs to the prompt.
So I replaced written rules with things that run on their own:
- A guard rail that blocks dangerous delete commands
- A size check that refuses a CLAUDE.md longer than 80 lines
- A forced checkpoint every 75 actions, so the work does not drift
- Two small helpers: one that reads large files, one that re-checks the work without trusting the author
- An evolve command that proposes an improvement, tests it, then keeps it or throws it away
The rule I keep
A rule you can check belongs in a script, not in a text.
The AI can forget or misread a text. It cannot ignore a script that blocks. My CLAUDE.md is shorter than in March, and the system is more reliable.