skill-issue – skill Agent & Tự động hóa cho agent AI

skill-issue — skill activation audit

Overview

A coding agent decides which skill to run from each skill's always-on name +

description. A skill can be perfectly implemented and still never fire because

its description is too vague to match how people phrase requests, or because a

more specific sibling silently wins. skill-issue audits exactly that surface,

grading each skill A–F, simulating which skill fires for a given prompt, and

reporting collision clusters where one skill shadows another.

When to Use This Skill

• Use when a skill you wrote never seems to trigger and you don't know why
• Use when the user says "why isn't my skill firing", "which skill fires for X", or "audit my skills"
• Use after writing or installing a new SKILL.md, to investigate likely activation issues
• Use in CI to fail a PR that adds a skill with empty/duplicate/colliding metadata

How It Works

Install the CLI (npm i -g @misha_misha/skill-issue, brew install mishanefedov/skill-issue/skill-issue, or npx @misha_misha/skill-issue), then:

skill-issue ~/.claude/skills                       # grade every skill A–F (+ collisions summary)
skill-issue ~/.codex/skills --why "deploy to prod" # which skill fires for this prompt, and why
skill-issue <dir> --collisions                     # clusters of skills that shadow each other
skill-issue <dir> --fix                            # append a "Use when …" clause to weak descriptions
skill-issue <dir> --json                           # machine-readable; exits non-zero on errors

Offline heuristic by default; add --llm to judge with a local claude/codex CLI.

Examples

Example 1: Audit installed skills

skill-issue ~/.claude/skills
# F  deploy-helper  ✗ no description — can never fire
# C  shipit         ! no "use when …" trigger clause
# A  rollback-prod  heuristic: trigger wording present; actual activation unverified

Example 2: Diagnose a collision

skill-issue ~/.claude/skills --why "deploy the app to prod"
#  1. shipit       0.74  ← would fire
#  2. land-deploy  0.69  (margin 0.05 — ambiguous, likely collision)

Limitations

• Offline scoring is heuristic and should be treated as a triage signal, not a final quality verdict.
• Collision reports highlight likely shadowing, but agent-specific routers can weight metadata differently.
• The --fix mode can improve weak trigger wording, but generated edits still need maintainer review before committing.

Execution boundaries

A score or simulated trigger is not proof that a native host activates a skill. Record actual client/version observations separately. Use --fix only within authorized paths after inspecting a proposed diff; preserve original metadata. --llm may send content through the configured model provider and needs the corresponding data/cost scope.