Skill Metrics & Evolution

Every skill has a lifecycle: creation, adoption, maturation, and eventually retirement. Metrics tell you where each skill is in its lifecycle and whether it is delivering value or consuming resources without return. The three metrics that matter are activation frequency, output acceptance rate, and correction density.

Activation frequency measures how often the skill fires. A skill that activates daily is high-value infrastructure. A skill that has not activated in 30 days is either solving a rare problem (valuable but dormant) or solving a problem nobody has anymore (candidate for retirement). Output acceptance rate measures how often the skill's output is used without modification. An 80% acceptance rate means the skill is well-calibrated. A 40% acceptance rate means the skill needs refinement — its output consistently requires human editing. Correction density measures how many corrections the skill has received over its lifetime. A skill with twenty corrections in its first month is actively evolving. A skill with zero corrections in six months is either perfect or abandoned.

Skill retirement is as important as skill creation. A retired skill is one that no longer provides enough value to justify its presence in the skill library. The retirement criteria are straightforward: zero activations in 90 days with no seasonal explanation, or an output acceptance rate below 30% that has not improved despite corrections. Retired skills are archived, not deleted — moved to a .claude/skills/archived/ directory where they consume no tokens but remain available if the need resurfaces.

The evolution pattern for healthy skills follows a predictable curve. Months 1-2: frequent corrections as edge cases are discovered. Months 3-4: corrections slow as the skill stabilizes. Months 5+: rare corrections, mostly driven by changes in the underlying tooling or requirements. If a skill is still receiving frequent corrections after month 4, it is either too broadly scoped (split it) or the underlying workflow is too unstable for automation (retire it and revisit later).