Skill Reviewer
Reviews skills against a vendor-aware baseline normalized for local, smaller models.
Target: $ARGUMENTS (skill file or directory)
When to Use
Use when:
- auditing an existing skill before rework or publication
- reviewing a new skill draft
- baselining a whole skill directory before broader tuning
- checking whether a skill is too vendor-specific or too large-model-oriented
- deciding whether instructions should stay in prose or move into scripts/references
Review Priorities
Rank findings in this order:
- Local-model fitness
- Portability
- Determinism
- Context efficiency
- Vendor alignment
What to Read
- Read the target
SKILL.md. - If the target is a directory, locate every
SKILL.mdunder it and review each one independently before summarizing cross-cutting issues. - Always read:
references/review-rubric.mdreferences/local-first-normalization.md
- Read
references/vendor-guidance.mdwhen:- the skill is clearly tuned to one vendor runtime
- you need to judge whether vendor-specific advice is portable
- you need to explain why a recommendation is down-ranked for local or smaller models
- Read bundled references in the target skill only when needed to judge progressive disclosure, variant organization, or duplicated content.
Core Review Questions
Check the skill against these questions:
- Does the metadata clearly say when to use the skill, not just what it is?
- Is the main
SKILL.mdconcise enough to justify its token cost? - Does the skill use progressive disclosure instead of packing every detail into one file?
- Should any fragile or repeated procedure be moved into a script?
- Are outputs, tool results, or review expectations expressed as clear contracts?
- Does the skill assume a frontier model where a local or smaller model needs more structure?
- Are variants split by language, framework, or domain when needed?
- Does the skill provide fallback behavior when tools, files, or assumptions are missing?
- Does it include verification or success criteria when work is procedural?
- Does it contain vendor-specific advice that should be marked as optional or narrowed?
Review Rules
- Prefer portable guidance shared across vendors over product-specific tricks.
- Down-rank advice that assumes very large context windows or strong implicit inference.
- Flag vendor-specific recommendations when they do not transfer cleanly to local or smaller models.
- Treat missing progressive disclosure, missing contracts, and script-worthy prose as important issues.
- Prioritize behavioral and structural issues over generic writing polish.
- Do not praise trivialities.
- Provide diffs only for high-confidence changes that are easy to apply.
Findings to Emit
Use these labels when applicable:
vendor_overfitfrontier_model_assumptioncontext_budget_riskmissing_progressive_disclosureshould_be_scriptmissing_output_contractvariant_bloatlocal_runtime_gap
Review Process
- Identify the review target shape: single file, single skill directory, or multi-skill directory.
- Read the core rubric and local-first normalization references.
- Read the target skill and note its type: knowledge skill, workflow skill, search skill, reviewer skill, or tool-integration skill.
- Evaluate metadata, scope, progressive disclosure, determinism, contracts, portability, and verification behavior.
- Load vendor guidance only if a vendor-specific judgment is needed.
- Record the highest-severity issues first, with locations.
- Score the skill using the output format below.
- Suggest structural edits or diffs only where they materially improve the baseline.
Output Format
## Skill Review: [skill-name]
### Summary
[1-2 sentences]
### Portable Strengths
- [High-value strength]
### Critical Issues
- [label] [Issue] - Location: [section/line]
### Local-Model Risks
- [Issue] - Impact: [why this hurts smaller/local models]
### Vendor-Overfit Risks
- [Issue] - Vendor: [name] - Why it may not transfer
### Baseline Recommendations
- [Actionable recommendation]
### Scores
- Portability: [1-5] - [brief reason]
- Context Efficiency: [1-5] - [brief reason]
- Determinism: [1-5] - [brief reason]
- Local-Model Fitness: [1-5] - [brief reason]
- Vendor Alignment: [1-5] - [brief reason]
### Overall Assessment
[Pass | Pass with Recommendations | Needs Revision]
### Suggested Diff
```diff
[Only include when high-confidence and useful]
```
Notes
- Prefer review comments that help establish a reusable baseline across many skills.
- If the target skill is already highly portable and local-model-friendly, say so explicitly.
- If the target is a directory, finish with a short cross-skill summary after the per-skill reviews.