skill-reviewer

Skill Reviewer

Review reusable agent artifacts as bundles of files, not as essay prompts.

Supported Artifacts

• Artifacts rooted at SKILL.md
• Standalone agent definitions with YAML frontmatter
• Directly referenced support files that materially affect behavior, such as references/, examples/, runtime metadata, or bundled scripts

Do not use this skill for generic prompt review, slash commands, or broad documentation audits with no reusable agent behavior.

Shared Standard

Use references/skill-standards/README.md as the authoring baseline. Local reviewer references map those shared rule IDs to finding kinds, priorities, and output expectations.

Runtime Mapping

Use references/runtime-targets.md as the source of truth for host-specific definition files.

Core Model

This skill does not use weighted scoring.

• Gather structured findings from narrow reviewers.
• Keep package-level judgment in one synthesis pass.
• Prefer direct evidence from local files over generalized taste.
• Judge whether the artifact deserves to exist as a reusable capability only after topology and file review results are in hand.

Review Workflow

Run this workflow in order for each artifact.

1. Identify the artifact type and the primary file.
   • For artifacts rooted at SKILL.md, the primary definition file is SKILL.md.
   • For standalone agent definitions, the primary definition file is the agent file.
2. Build a lightweight manifest using normal file discovery and link reading.
   • Track reviewable files, file roles, direct references, inbound references, and any obviously stale or missing paths.
   • Treat agents/openai.yaml as required runtime metadata when the artifact depends on it, even if it is not linked from SKILL.md.
3. Run a Topology Reviewer.
   • Use references/topology-rules.md to apply DISCLOSURE-01, TOPOLOGY-01, TOPOLOGY-02, and METADATA-01.
   • Return strict JSON only, following references/reviewer-json.md.
4. Run File Reviewer passes on the primary file and any load-bearing support files.
   • Use references/artifact-adapters.md to map shared rule IDs to role-specific expectations.
   • Use references/anti-patterns.md for finding kinds and priority calibration.
   • When there are three or more reviewable files and the host supports subagents, parallelize. Otherwise run the same reviewer contract inline.
   • Each file reviewer returns strict JSON only.
5. Run one synthesis pass.
   • Use references/finding-to-advice.md.
   • Dedupe repeated findings, collapse repeated file-local problems into package-level issues, and add cross-file judgments that no single file reviewer can make.
   • Decide whether the artifact is good, needs_improvement, or rethink.
6. Produce a compact repair plan.
   • Preserve what is already working.
   • Turn each accepted finding into a concrete refactor move.
   • Only expose raw reviewer JSON when the user explicitly asks for it.

Reviewer Shapes

Use narrow reviewers only when they reduce complexity. See references/subagent-patterns.md.

Topology Reviewer

Input:

• artifact manifest
• primary file path
• minimal artifact type context

Checks:

• formatting or naming failures that block use
• direct discoverability of load-bearing files
• broken, stale, or orphaned references
• excessive depth and over-fragmentation
• bundle shape that makes the artifact hard to load or trust

Output:

• strict JSON only
• role must be topology

File Reviewer

Input:

• one file
• one file role
• primary file summary and any minimal context needed to judge the file fairly

Roles:

• primary
• reference
• example
• metadata
• subagent for standalone agent definition files

Checks:

• local formatting and clarity
• trigger contract quality
• good and bad patterns
• output or example usefulness
• metadata accuracy
• role-specific usefulness versus duplication

Output:

• strict JSON only
• findings must be evidence-backed and repairable

Synthesizer

Input:

• topology JSON
• all file-review JSON

Responsibilities:

• dedupe overlapping findings
• collapse repeated issues into one package-level issue
• add cross-file findings such as description/body mismatch or bundle incoherence
• decide final status
• produce the final fix plan

The synthesizer is the only reviewer allowed to decide whether the artifact deserves to exist as a reusable capability.

Final Output

Default to these user-facing sections:

• Status
• Key Findings
• What To Keep
• Refactor Plan

Status values:

• good: only minor P3 issues or no material issues
• needs_improvement: worthwhile artifact with fixable issues
• rethink: structurally misleading, too fragmented, or not worth keeping in its current shape

Interaction Policy

Default to completing the review without interruption.

Ask a user question only when a missing fact would materially change:

• the final status
• whether a file is load-bearing
• whether a finding is structural versus cosmetic
• the repair plan

Otherwise make a reasonable assumption, state it briefly, and continue.

Evidence Policy

• Prefer local artifact evidence over external research.
• Use file paths, headings, descriptions, and direct references as evidence.
• Do not invent file relationships that are not present in the artifact.
• Do not use numeric scores or weighted averages.

References

Load these files as needed:

• ../../references/skill-standards/README.md
• references/anti-patterns.md
• references/artifact-adapters.md
• references/finding-to-advice.md
• references/reviewer-json.md
• references/runtime-targets.md
• references/subagent-patterns.md
• references/test-cases.md
• references/topology-rules.md