I have been writing cards for a while now, little markdown files the agent loads only when the situation matches. And the setup procedure I keep on ai.memention.net happily produces ten or twelve of them for a repo. Architecture card, domain cards, feature cards. But then I was looking at a small Flutter tool for editing translation strings and thought, why am I explaining go_router to Claude? It already knows go_router. Better than me, probably.

So how about just naming it? One line per thing the model already knows, plus whatever is odd about how it is used here:

- **go_router** — all routes in `lib/app/router.dart`; pages get their dependencies via constructor args there.
- **Dialogs** — each dialog is a widget with a `static Future<void> show({context, …, onX})`

And cards only for the stuff it can’t know. My own widget DSL that isn’t on pub.dev, a sed hack that toggles web-only lines, and the data model that lives on a PocketBase server outside the repo. Three cards instead of twelve.

But frankly, a gut feeling is not a result. So I made four copies of the repo, no docs, cards only, cards plus the known stack lines, and the full twelve card set, and let Sonnet do three tasks in each, three times. 36 runs, $4.44, and a blind judge scoring the answers.

The known stack copy needed 4.9 turns on average. The same cards without those few lines needed 8.2. About 1.6 KB of text, and the agent skipped half the poking around. The full card set cost about twice as much on the code tasks and scored lower, mostly for wandering off and doing things nobody asked for. All that architecture prose said nothing the code didn’t already say. The known parts already exist as a kind of intuition in the model, one only need to wake them up. A name does that. Me smile.

And the copy with no docs at all? It did ok too. Slower and a bit sloppier, but ok. That gave me a feeling I had back in the late 90s, when I worked on simulators with Silicon Graphics computers. They had so much compute power that even bad programs looked good. The models are a bit like that now, one can give them pretty bad prompts and still get kind of good results back. But with a little setup one get even better results, faster and cheaper.

It is a 700 line repo and one model, so who knows how it holds up on something bigger. It is in the setup now as an experimental pattern anyway, grab it at ai.memention.net/cards/known-stack.md…