Skill File Validator

Validate skill files for correct format against the official Claude Code spec.

Validation Target

• With argument: validate specific file
• Without: validate all .claude/skills/*/SKILL.md

Frontmatter Reference (official fields)

Skills support these fields. Only description is meaningfully required (a skill with none of these still loads — name defaults to the directory name).

---
name: skill-name              # Optional: defaults to directory name; lowercase, hyphenated
description: What it does      # Recommended: shown in / menu and used for triggering
when_to_use: When to reach for it  # Optional: appended to description for triggering
                              #   (description + when_to_use truncated at 1,536 chars in listings)
user-invocable: true          # Optional: default true; false hides from / menu (Skill tool still works)
model: sonnet                 # Optional: any /model value (sonnet, opus, haiku, fable, best,
                              #   default, opusplan, sonnet[1m], opus[1m], a full model ID) or inherit
effort: medium                # Optional: low, medium, high, xhigh, max
allowed-tools:                # Optional: pre-approved tools. Accepts a comma-separated string
  - Read                      #   OR a YAML list — BOTH are official (list shown here as house style)
  - Grep
  - Glob
disallowed-tools:             # Optional: explicitly blocked tools (string or YAML list)
  - WebFetch
context: fork                 # Optional: run the skill in a subagent
agent: agent-name             # Optional: subagent type when context: fork (default: general-purpose)
background: false             # Optional (2.1.218+): with context: fork, false waits for the result;
                              #   forked skills run in the background by default
argument-hint: "[issue-number]"  # Optional: autocomplete hint shown after the skill name
arguments: target flags       # Optional: space-separated argument NAMES for $name substitution
                              #   (e.g. arguments "target flags" -> use $target and $flags in the body).
                              #   NOT a structured array of name/description/required objects.
disable-model-invocation: false  # Optional: bool; true also blocks scheduled-task invocation
paths:                        # Optional: glob patterns that limit where the skill activates
  - "src/**/*.ts"
shell: bash                   # Optional: bash or powershell for inline !`command` blocks
hooks:                        # Optional: skill-scoped lifecycle hooks (same schema as settings.json)
  PreToolUse:
    - matcher: Write
      hooks:
        - type: command
          command: ./scripts/validate.sh
  Stop:
    - hooks:
        - type: command
          command: ./scripts/verify.sh
          once: true
---

Valid Tools (for allowed-tools / disallowed-tools)

Read, Write, Edit, Bash, Grep, Glob, Agent (Task = legacy alias),
Skill, WebFetch, WebSearch, TodoWrite, NotebookEdit, AskUserQuestion

Non-official fields (flag as warnings, not errors)

metadata, license, version   # Not part of the skill spec; harmless but should be removed
once                         # Valid ONLY inside a hooks[] entry, never at skill top level

Validation Checklist

Required / Recommended Fields

• description present (recommended; used for the / menu and triggering)
• name, if present, is lowercase and hyphenated (defaults to directory name if omitted)

Optional Field Validation

• allowed-tools / disallowed-tools are valid tool names
• allowed-tools / disallowed-tools are a comma-separated string OR a YAML list (both valid)
• allowed-tools does not broadly preapprove bare Bash, Edit, Write, Agent, Task, or Skill; omit it when normal runtime approval is required
• model is a /model value (fable, opus, sonnet, haiku, best, default, opusplan, sonnet[1m], opus[1m], a full model ID) or inherit
• effort is one of: low, medium, high, xhigh, max
• context is fork (if specified)
• agent only set alongside context: fork (defaults to general-purpose)
• background is boolean and only set alongside context: fork (default true since 2.1.218)
• argument-hint is a string
• arguments is a space-separated string of argument names (NOT a structured array)
• when_to_use is a descriptive string (appended to description for triggering)
• description + when_to_use stay within 1,536 characters (the listing truncates beyond it)
• disable-model-invocation is boolean
• paths is a list of glob patterns
• shell is bash or powershell
• user-invocable is boolean
• hooks has valid structure (matcher + hooks[] with type/command)

Unknown / Non-official Fields

• Accept metadata (map), license, and compatibility (<=500 chars): official but not acted on
• Warn on version and any other field outside the official list; Claude Code ignores them silently
• Warn on once used at the top level (only valid inside a hooks[] entry)

Content Structure

• Clear instructions
• Uses $ARGUMENTS (or $name per arguments) if expecting input
• Step-by-step process if complex

Output Format

## Skill Validation Report

### Files Checked
| File | Status | Issues |
|------|--------|--------|
| workflow/SKILL.md | OK | None |
| my-skill/SKILL.md | WARN | Missing description |

### Summary
- Total: [N]
- Valid: [N]
- Needs fixes: [N]

Auto-Fix

• Add missing name from directory name
• Add missing description
• Normalize a comma-separated allowed-tools string to a YAML list (house style only — the string form is valid; do NOT flag it as an error)
• Remove non-official fields (metadata, license, version)
• Convert a structured arguments array to a space-separated name string
• Replace opusplan model with a supported value
• Add $ARGUMENTS handling