Improve IC Skill

A token-efficient workflow for improving an existing skill in the IC Skills repo. The guiding principle: understand the problem first, change precisely, verify only what changed. Evals exist as regression safety nets — run them selectively, not automatically on every pass.

Step 0 — Understand the problem before touching anything

This is a hard stop. Do not read any file, run any command, or make any change until you have a specific problem statement from the user.

You need to know:

1. What is the specific issue? (e.g. "canister ID X is stale", "missing pitfall about Y", "description doesn't trigger on Z", "code example uses deprecated API W")
2. Why does it matter? (e.g. "agents generate broken code", "skill never triggers for this use case", "upstream sync added a new command we don't cover")

If the request is vague ("improve the motoko skill", "the ckbtc skill feels weak") — ask:

"What specific issue should I focus on? For example: a missing pitfall, a stale canister ID, a broken code example, or a description that isn't triggering correctly."

If the user cannot or will not provide a specific issue — stop. Do not proceed. Do not read the skill. Do not run commands. Do not "explore and find improvements yourself." Explain clearly:

"I need a specific problem to solve before I can start. Blind improvements risk changing things that don't need changing. To find a concrete issue, try: running npm run validate, checking evaluations/<skill-name>.json for failing cases, or pointing me to a GitHub issue or upstream diff."