foundation

Every skill in this repository inherits the rules below. The root virtue is predictability. The agent takes the same process every run. A skill earns its place by removing ambiguity.

The determinism ladder

Climb until a rung holds, then stop:

1. A script or existing tool can do the work. Call it. The output is deterministic and free of agent judgment.
2. A linter or schema can validate the work. Gate on it.
3. A deterministic primitive exists for the operation. Use skillkit instead of improvising the file operation in prose.
4. Only the irreducibly stochastic part is left. Hand that part to the model, and nothing else.

Push as much work down the ladder as the task allows. The model is the last resort.

Determinism is token-efficiency

Pushing work down the ladder also cuts tokens. The two goals never conflict. A script deletes output the model would otherwise generate token by token, then hands back a verdict the agent reads in one line. The agent branches on the exit code and reads the artifact back only when the script failed. Success is quiet by design: a tool prints one line or nothing, because verbose stdout the agent re-reads spends the tokens the script just saved. The model is authored into the tool once. After that it runs for free.

Non-negotiable rules