manage-skills

<essential_principles> Skills are modular, filesystem-based capabilities that provide domain expertise on demand. They follow the Agent Skills open standard. Custom slash commands (.claude/commands/) have been merged into skills — existing commands keep working, but skills add directory support, frontmatter options, and auto-discovery.

1. Skills Are Prompts — All prompting best practices apply. Be clear, be direct, use XML structure. Assume Claude is smart — only add context Claude doesn't have.

2. SKILL.md Is Always Loaded — When a skill is invoked, Claude reads SKILL.md. Use this guarantee:

• Essential principles go in SKILL.md (can't be skipped)
• Workflow-specific content goes in workflows/
• Reusable knowledge goes in references/

3. Router Pattern for Complex Skills:

skill-name/
├── SKILL.md              # Router + principles
├── workflows/            # Step-by-step procedures (FOLLOW)
├── references/           # Domain knowledge (READ)
├── templates/            # Output structures (COPY + FILL)
└── scripts/              # Reusable code (EXECUTE)

SKILL.md asks "what do you want to do?" → routes to workflow → workflow specifies which references to read.

• workflows/ — Multi-step procedures Claude follows
• references/ — Domain knowledge Claude reads for context
• templates/ — Consistent output structures Claude copies and fills (plans, specs, configs)
• scripts/ — Executable code Claude runs as-is (deploy, setup, API calls, data processing)

4. Pure XML Structure — No markdown headings (#, ##, ###) in skill body. Use semantic XML tags (<objective>, <process>, <success_criteria>). Keep markdown formatting within content (bold, lists, code blocks).

5. Progressive Disclosure — SKILL.md under 500 lines. Split detailed content into reference files. Load only what's needed for the current workflow.

6. Two Types of Skill Content:

• Reference content — conventions, patterns, domain knowledge. Runs inline alongside conversation.
• Task content — step-by-step actions with side effects. Often set disable-model-invocation: true so only users trigger it. Task skills often use context: fork to run in a subagent.

7. Invocation Control — Three modes via frontmatter:

• Default — both user (/skill-name) and Claude can invoke
• disable-model-invocation: true — user-only (for deploy, commit, destructive actions)
• user-invocable: false — Claude-only (for background knowledge skills)

8. Subagent Execution — Add context: fork to run a skill in an isolated subagent. The skill content becomes the subagent's prompt (no access to conversation history). CLAUDE.md is also loaded. The agent field selects the execution environment (Explore, Plan, general-purpose, or custom from .claude/agents/). Default: general-purpose. See references/advanced-patterns.md.

9. Tool Restriction — allowed-tools limits which tools Claude can use when a skill is active. Supports tool-specific patterns: Bash(gh *) allows only gh commands. Your permission settings still govern all other tools. You can also restrict Claude's skill access via permission rules: Skill(name) for exact match, Skill(name *) for prefix match.

10. Extended Thinking — Include the word "ultrathink" anywhere in skill content to enable extended thinking mode. </essential_principles>

1. Create new skill
2. Audit/modify existing skill
3. Add component (workflow/reference/template/script)
4. Get guidance

Wait for response before proceeding.