Claude Skill

skill-health-audit

Audit an agent skill or skillpack for trigger fit, false activation, structure, metadata, intended use, risks, composability, model sensitivity, fixtures, held-out evidence, prompt overhead, catalog drift, and regression risk. Use when deciding whether a skill is ready to publish

LLM Mart · 0 points · 2 views 0 listing impressions 0 install-command copies
Download markoblogo-abvx-agent-skills-skills_skill-health-audit-37a7146.zip · 3 KB

Install

skills CLI npx skills add https://github.com/markoblogo/abvx-agent-skills/tree/main/skills/skill-health-audit
Claude Code claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install markoblogo-abvx-agent-skills@llmmart
Git git clone https://github.com/markoblogo/abvx-agent-skills.git

The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole markoblogo/abvx-agent-skills collection as a plugin from our marketplace. Git is the plain clone.

Skill manifest

Skill Health Audit

Review a skill as a versioned procedural artifact. The audit produces evidence for a maintainer; it does not become an autonomous editing or publishing loop.

Audit Lenses

Check the target skill and its neighboring public surfaces for:

  • trigger fit and false activation;
  • required structure and metadata;
  • intended use, out-of-scope boundaries, risks, and anti-patterns;
  • composability and conflict with neighboring skills;
  • model sensitivity and instruction-following burden;
  • activation cases, positive fixtures, negative cases, and fixture/held-out evidence;
  • prompt bloat, repeated rules, and context overhead;
  • catalog/README drift, card drift, and generated-artifact drift;
  • regression risk, unsafe authority expansion, and missing human gates;
  • the highest evidence-supported evaluation tier.

Method

  1. Identify the target skill version, source paths, and evaluation scope.
  2. Read the trigger and output contract before reading supporting prose.
  3. Inspect the card, metadata, companion skills, fixtures, and catalog entry.
  4. Run only the available deterministic checks; label absent runtime evidence.
  5. Check one positive activation and one negative/anti-pattern case where fixtures exist.
  6. Rank findings by user harm, false confidence, drift, and maintenance cost.
  7. Recommend structural_only, fixture_checked, rollout_checked, or held_out_validated only when the captured evidence supports it.

Review Status

Return exactly one recommendation:

  • ACCEPT — contract and evidence are adequate for the stated tier;
  • REVISE — concrete changes are needed before acceptance;
  • INSUFFICIENT_EVIDENCE — the skill may be sound but the claimed tier is not proven;
  • REJECT — the skill is unsafe, redundant, misleading, or not reusable.

Output

Skill: <name>
Version: <card/frontmatter version>
Recommended tier: <tier>
Status: ACCEPT | REVISE | INSUFFICIENT_EVIDENCE | REJECT

Trigger fit: <finding>
Structure/metadata: <finding>
Evidence: <fixtures, reports, or missing proof>
Drift: <catalog/README/card result>
Regression and authority risk: <finding>
Top findings: <ranked list>
Maintainer decision needed: <yes/no and scope>

The audit must not edit skills, catalog files, best_skill.md, release files, or installed skills. It must not promote a draft or treat a persuasive prose review as behavioral evidence.

Composition

  • Use bounded-evaluation to design stronger fixture or held-out checks.
  • Use private-vs-publishable-skill-audit before public extraction.
  • Use context-degradation-review for prompt bloat or instruction conflict.
  • Use confidence-fragility-review when the claimed tier exceeds its proof.
Files (abvx-agent-skills)
  • agents
    • openai.yaml 284 B
      Loading…
  • SKILL.md 3.1 KB
    ---
    name: skill-health-audit
    description: Audit an agent skill or skillpack for trigger fit, false activation, structure, metadata, intended use, risks, composability, model sensitivity, fixtures, held-out evidence, prompt overhead, catalog drift, and regression risk. Use when deciding whether a skill is ready to publish, revise, evaluate, or keep structural-only.
    license: MIT
    metadata:
      abvx_status: experimental
      abvx_origin: original
      abvx_eval_tier: structural_only
    ---
    
    # Skill Health Audit
    
    Review a skill as a versioned procedural artifact. The audit produces evidence
    for a maintainer; it does not become an autonomous editing or publishing loop.
    
    ## Audit Lenses
    
    Check the target skill and its neighboring public surfaces for:
    
    - trigger fit and false activation;
    - required structure and metadata;
    - intended use, out-of-scope boundaries, risks, and anti-patterns;
    - composability and conflict with neighboring skills;
    - model sensitivity and instruction-following burden;
    - activation cases, positive fixtures, negative cases, and fixture/held-out evidence;
    - prompt bloat, repeated rules, and context overhead;
    - catalog/README drift, card drift, and generated-artifact drift;
    - regression risk, unsafe authority expansion, and missing human gates;
    - the highest evidence-supported evaluation tier.
    
    ## Method
    
    1. Identify the target skill version, source paths, and evaluation scope.
    2. Read the trigger and output contract before reading supporting prose.
    3. Inspect the card, metadata, companion skills, fixtures, and catalog entry.
    4. Run only the available deterministic checks; label absent runtime evidence.
    5. Check one positive activation and one negative/anti-pattern case where
       fixtures exist.
    6. Rank findings by user harm, false confidence, drift, and maintenance cost.
    7. Recommend `structural_only`, `fixture_checked`, `rollout_checked`, or
       `held_out_validated` only when the captured evidence supports it.
    
    ## Review Status
    
    Return exactly one recommendation:
    
    - `ACCEPT` — contract and evidence are adequate for the stated tier;
    - `REVISE` — concrete changes are needed before acceptance;
    - `INSUFFICIENT_EVIDENCE` — the skill may be sound but the claimed tier is not
      proven;
    - `REJECT` — the skill is unsafe, redundant, misleading, or not reusable.
    
    ## Output
    
    ```text
    Skill: <name>
    Version: <card/frontmatter version>
    Recommended tier: <tier>
    Status: ACCEPT | REVISE | INSUFFICIENT_EVIDENCE | REJECT
    
    Trigger fit: <finding>
    Structure/metadata: <finding>
    Evidence: <fixtures, reports, or missing proof>
    Drift: <catalog/README/card result>
    Regression and authority risk: <finding>
    Top findings: <ranked list>
    Maintainer decision needed: <yes/no and scope>
    ```
    
    The audit must not edit skills, catalog files, `best_skill.md`, release files,
    or installed skills. It must not promote a draft or treat a persuasive prose
    review as behavioral evidence.
    
    ## Composition
    
    - Use `bounded-evaluation` to design stronger fixture or held-out checks.
    - Use `private-vs-publishable-skill-audit` before public extraction.
    - Use `context-degradation-review` for prompt bloat or instruction conflict.
    - Use `confidence-fragility-review` when the claimed tier exceeds its proof.
    
  • SKILL_CARD.md 2.2 KB
    Loading…

Comments (0)

Sign in to join the conversation.

No comments yet.

Reviews (0)

No reviews yet.