/skill-review

Review Criteria Reference

Detailed criteria for each review dimension. Load this file during Step 2 analysis.

Structure Compliance — Detailed Checks

Required

• SKILL.md exists at root
• Frontmatter has name field (matches directory name)
• Frontmatter has description field (non-empty, >50 chars)

Expected

• SKILL.md ≤ 500 lines (hard limit; flag if >400 as approaching)
• No files that serve no purpose: README.md (unless skill IS about READMEs), CHANGELOG.md, CONTRIBUTING.md
• scripts/ — if present, all files are executable and tested
• references/ — if present, each file is referenced from SKILL.md with load instructions
• assets/ — if present, files are used in output, not loaded into context

Red Flags

• SKILL.md > 500 lines → must extract to references
• Deeply nested references (2+ levels) → flatten
• Orphaned files (not referenced from SKILL.md)
• Scripts without clear invocation instructions in SKILL.md

Description Quality — Detailed Checks

Keyword Bombing Indicators (Good)

• 3+ verb variations for the same action (create/build/write/make)
• Noun variations (skill/agent/tool/plugin)
• Natural phrases a user would type
• Both imperative ("create a skill") and question form ("how do I build a skill")

Description Anti-Patterns (Bad)

• Single-sentence description under 100 characters
• Only technical terms, no natural language triggers
• Describes implementation ("Uses markdown frontmatter...") instead of user intent
• Contains instructions meant for the model body

Scoring Heuristic

• Count unique trigger phrases → <3 is weak, 5-8 is good, >10 is excellent
• Check if casual user language would match → try 3 hypothetical queries mentally

Workflow Design — Detailed Checks

Checklist Quality

• Uses - [ ] format (copy-paste friendly)
• Steps are numbered and sequential
• Sub-steps use indentation
• ⚠️ marks appear on steps that must not be skipped
• ⛔ marks appear on prerequisites/blockers

Confirmation Gates

Must exist before:

• File creation/deletion/overwrite
• External API calls with side effects
• Long-running generative operations
• Applying analysis results to code

Flow Quality

• Can a reader understand the full process from the checklist alone?
• Are conditional branches clearly marked?
• Is there a clear start and end state?

Token Efficiency — Detailed Checks

Iron Law

• Present: yes/no
• Placed at top (before workflow): yes/no
• Specific and falsifiable: yes/no (bad: "write good code"; good: "never use placeholder text")
• Addresses the #1 likely failure mode: yes/no

Progressive Loading Signals (Good)

• "Load references/X.md for..." with clear trigger conditions
• References loaded at specific workflow steps, not all in Step 1
• Large context only loaded when that branch is taken

Bloat Signals (Bad)

• Explaining things Claude knows (how to write markdown, what JSON is)
• "You should" / "You will" / "Remember to" (wasted tokens on filler)
• Duplicated instructions (same thing said in workflow AND in later section)
• Comments/explanations aimed at human readers, not model behavior

Anti-Pattern Detection — Detailed Examples

Vague Directives (rewrite as questions)

• ❌ "Ensure the output is high quality"
• ✅ "Does every heading have ≥2 sentences of content beneath it?"

Over-Specification

• ❌ "Use markdown headers with # for h1, ## for h2..." (Claude knows this)
• ✅ "Use exactly 2 levels of headers: # for sections, ## for subsections" (this constrains)

Missing Guardrails

• No anti-patterns section = model will take lazy defaults
• No pre-delivery checklist = no verification step
• No Iron Law = no north star for quality decisions

Monolithic SKILL.md

• If SKILL.md > 300 lines, check: could any section be a reference loaded on-demand?
• Domain knowledge that applies to only one step → extract to reference
• Examples longer than 10 lines → extract to reference