agent-ebpf skill-usage-insights.instructions.md

skill-usage-insights

Skill Usage Insights

Turn .claude/learning/runs.jsonl (the [[self-learning]] run log) and the contents of .claude/skills/ into a per-skill usage/KPI report, with add/remove recommendations.

1. Gather inputs

• .claude/skills/<name>/SKILL.md — skills currently installed in this project (one directory per skill).
• .claude/learning/runs.jsonl — append-only JSON-lines log; each line: {"ts": ..., "skill": ..., "action": ..., "rc": 0|nonzero, "duration": ..., "error": "", "hint": "", "note": ""}. If this file doesn't exist, say so plainly — there's no usage history yet, not an error. Recommendations in that case are limited to "not yet measurable; check back after the self-learning skill has recorded some runs".
• .claude/learning/skill-feedback.jsonl — user negative reactions per skill (written by [[skill-feedback-adaptation]]). Use for inefficiency % and update suggestions alongside run-based KPIs.
• Optionally, .claude/learning/task-skill-proposals.json — latest task-scoped skill proposals (also from [[skill-feedback-adaptation]]).
• Optionally, the library's manifest.json (wherever the skill library lives, e.g. ~/.claude/skills/ or a synced skills_library/) — useful to see a skill's detect_globs/description when judging relevance.

2. Aggregate per skill

For each installed skill name, and for each distinct skill value in runs.jsonl that matches an installed skill name:

• runs — count of matching records.
• success_rate — rc == 0 count / runs (as a percentage).
• avg_duration — mean of duration across records that have it.
• last_used — max ts; days_since_last_use — days between that and now.
• Per agent — when agent is set on a row (claude, cursor, kiro, copilot), group invocations by agent. The VS Code Usage Report shows a Skill usage by agent section and a matrix for skills used by more than one agent on the same workspace (common when switching between Claude Code, Cursor, Kiro, and Copilot on one task). Rows without agent count as Claude.
• Records whose skill value doesn't match any installed skill name are general task-tracking entries (per [[self-learning]]'s schema) — ignore them for this report, but you may mention the count as "N other tracked task runs not tied to a specific skill".