vet-skill-idea

Every skill's description loads into context for every session in its scope. An unworthy skill is permanent tax that also dilutes activation for the skills that earn their place — so the default answer is no. The job here is to find the reason not to build a skill; only an idea that survives every gate below deserves one.

The gates

An idea warrants a skill only by clearing all four. Each gate carries the test that discriminates a pass from a failure, and where a failure routes instead.

If the idea description is too vague to test against a gate, ask one focused question to clarify it before proceeding — do not evaluate an idea you cannot fairly test.

Net-new capability

A skill is worth its context only if the agent does the task materially worse without it. Test against the base model: would a competent agent with no skill produce a worse result? If it already does the job well, the skill teaches nothing it doesn't already know — drop it. Restating knowledge the agent carries from training is the most common way this gate fails.

Sound goal, not a gamed metric

A skill should instill an idea that moves the work toward the real outcome every time it runs — not chase a metric for its own sake. Test: can the skill be satisfied while missing the outcome? If success is box-ticking rather than the outcome landing, the skill optimises the metric and forgets the outcome. Reframe it around the outcome so that using it causes the outcome, or drop it.

Single responsibility

One unified job. Test: state the purpose in one sentence without "and", and write a description that fires on one clear class of situations. An idea that needs several unrelated triggers is several skills wearing one name — split it into those pieces and run each back through the gates.