Alpha Version: You are viewing the ALPHA documentation. This is an experimental version and may contain breaking changes.
Skip to main content

Testing Skills

Verification Checklist

Before shipping a skill, verify:

  1. Frontmatter validityname and description are present and within limits
  2. Reference file existence — every file mentioned in SKILL.md exists
  3. No tool-specific language — no "Bash tool", "Read tool", "Agent tool" in skill body
  4. Activation test — the skill triggers on relevant prompts
  5. Cross-assistant test — verify the skill works in at least one non-Claude assistant

Automated Checks

# Validate JSON config files
cat .claude-plugin/marketplace.json | python3 -m json.tool > /dev/null

# Check for Claude-specific tool names in skills
grep -rn "Bash tool\|Agent tool\|Read tool\|Write tool" .claude/skills/

# Verify all referenced files exist
for skill_dir in .claude/skills/*/; do
grep "references/" "$skill_dir/SKILL.md" | grep -oE 'references/[a-z-]+\.md' | while read ref; do
[ -f "$skill_dir$ref" ] || echo "MISSING: $skill_dir$ref"
done
done

Testing Activation

Start a new session (skills are discovered at session start), then test:

  1. Ask a question that should trigger the skill
  2. Verify the skill is loaded (check the system prompt for the skill description)
  3. Verify reference files are accessible

Cross-Assistant Testing

AssistantHow to Test
CursorOpen the project — .claude/skills/ is cross-read automatically
GitHub CopilotOpen in VS Code with Copilot agent mode
Codex CLICopy skills to .codex/skills/ or .agents/skills/

Using /validate-skill

If the workshop plugin is installed, use the AgentSkills spec validator:

/validate-skill .claude/skills/rescript