skills-auditlisted

Use to evaluate Agent Skills and detect when one is failing — a static health audit (empty/first-person/over-budget descriptions, missing triggers, near-duplicates, misplaced) plus an on-demand runtime trigger-eval (run a realistic prompt with the skill enabled vs disabled, 3x, score false positives/negatives). Classifies failures (under-triggered, over-triggered, wrong output, ignored, obsolete) and, behind a human gate, files a SANITIZED issue to the repo that OWNS the skill — the tool's own board for its skills, the project's board for the project's own skills, local-only for third-party — never leaking the private project you are working in. Triggers — "audita mis skills", "esta skill no dispara/falla", "evalúa el triggering", "skill health check", "por qué no se activó la skill", /skills audit.
CSalcedoDataBI/agentic-board · ★ 0 · AI & Automation · score 69

Install: claude install-skill CSalcedoDataBI/agentic-board

# skills-audit — detect skill failures, file them where they belong Part of the **skills-ops** module. Reuses `gh-account` (token), the `abios-feedback` sanitize-then-file discipline, `projects-admin` (issue + board), and the `guard-no-private.ps1` backstop. See `references/filing.md` for the filing recipe. ## Step 1 — Static audit (deterministic) ```powershell $audit = & "${CLAUDE_PLUGIN_ROOT}/scripts/Invoke-SkillAudit.ps1" -Root . -Scope all $audit.summary; $audit.findings | Format-Table severity,type,skill,filing ``` Findings carry a `filing` route (`file` = open an issue on the owner repo; `local` = report only). This is `Resolve-SkillOwner` deciding where each finding belongs — the tool's board for agentic-board skills, the project's board for project skills, local-only for third-party/personal (never open issues in someone else's repo). ## Step 2 — Runtime trigger-eval (on-demand, agentic) The static pass cannot tell if a skill actually *fires*. For a suspect skill, run the enabled-vs-disabled baseline — the core test: 1. Pick 3+ realistic prompts that SHOULD trigger the skill (and a couple that should NOT). 2. Toggle the skill with `skillOverrides` in `.claude/settings.local.json` (`"skill-name": "off"` to disable, remove to enable). Run each prompt **3×** per state. 3. Score: - **under-triggered** — enabled state shows no behavior change / you had to invoke it by hand. - **over-triggered** — fires on the should-NOT prompts (false positive). - **wron