I’ve been building an AI sound engineer. It runs in Reaper DAW, and it’s still nowhere near done, which is maybe the important part :)
The hard part early on wasn’t really audio. It was more the annoying, persistent problem of keeping Claude Code pointed in the right direction across sessions, which sounds small until it isn’t.
I made a claude.md and put the rules in it, and then every session I’d add a few more. Don’t touch the migrations. Run the tests first. Use this naming and not that one. Little things, mostly. The file kept growing, which felt fine for a while.
Then at some point Claude started ignoring rules it used to follow. Ones I’d written carefully. They were still in the file, technically. It just wasn’t acting on them anymore.
And the annoying bit was that the file had gotten too long to be read properly when a session started. Claude.md loads into context as the session opens, and context is a budget, so the file is competing with everything else for space.After a certain size, the rules near the bottom just stop landing. They’re there in the technical sense. They just don’t get followed. And the longer the file got, the worse that got.
So the fix was boring, which I think is usually a good sign. Trim it, restructure it, test it next session, repeat. I read a lot of docs to get the shape right, mostly Anthropic’s own. The part I didn’t expect was that I wasn’t really choosing the structure by taste. Claude Code and the docs kept nudging me toward the same shape, and what finally worked just happened to match what Anthropic documents. So it wasn’t really my preference, more like the pattern that was already there the whole time.
And the way to make the file shorter wasn’t to slash it randomly. It was to ask the same question of every line, over and over.
Can a machine check this? Then it probably doesn’t belong in prose. “Never edit /migrations/” is not something Claude should have to remember; that’s the kind of thing a hook should enforce every time, without any judgment.
Is it real judgment? Then keep it. “Match the existing style in the file you’re editing” can’t really be a hook, because the model has to look at it and decide. That’s prose territory.
Is it just filler, or the same thing said twice? Then cut it. “Write clean code” doesn’t really add anything. I had the same test rule written twice in slightly different words, and one of them just had to go.
You point it at your claude.md and it labels each line; move to a hook, keep as prose, or cut. It runs offline, so there’s no model call, which means the same file always gives the same result: https://github.com/blacksundev/claudemd-audit. Apache-2.0.
Run it on your own claude.md and tell me where it’s wrong. The classifier is heuristic, so of course it will mislabel some lines, and honestly that’s the part I care about most.
claudemd-audit is free and stays free. If you want the hooks and the tight claude.md templates I landed on, there's a pack -https://blacksundev.gumroad.com/l/ylbgvm , €19. You don't need it to use the tool.
Top comments (0)