Skill Creator

Skill Creator

Create high-quality Claude Code skills following progressive disclosure — metadata always loaded, SKILL.md body on trigger, references on demand.

Core Process

1. Interview

Gather requirements through focused questions. Limit to 3-4 AskUserQuestion calls maximum:

1. What should this skill enable?
2. When should it trigger? (user phrases, file patterns, contexts)
3. What tools does it need? (Read, Write, Edit, Bash, Agent, etc.)
4. Where should it run? (Skills loaded into the main session inherit the user's chosen model. Only specify a model when the skill ships an agent with its own override.)

Skip questions where the answer is obvious from context.

2. Explore Existing Patterns

Use Glob + Read directly to find existing conventions in the target project:

• Glob **/skills/**/SKILL.md to list defined skills, read 2-3 closest matches
• Sample the file structure and naming conventions via ls
• Detect the technology stack from manifest files (package.json, composer.json, pubspec.yaml)

3. Dedup Audit

Read ${CLAUDE_SKILL_DIR}/../prompt-writer/references/cc-dedup-guide.md before drafting. Check what CC already provides natively — the skill body must not repeat general concepts the model already knows. Cut anything CC handles by default.

4. Draft Skill

Follow progressive disclosure. SKILL.md body ≤500 lines. Move detail to references/. Read references/skill-patterns.md for description patterns, invocation control, string substitutions, and directory structure before writing.

Key decisions at this stage:

• Which description pattern fits? (A/B/C/D — see skill-patterns.md)
• user-invocable: false? disable-model-invocation: true? (see Invocation Control Matrix)
• context: fork if the skill is self-contained and parallelizable?
• paths: if the skill activates only for specific file types?

5. Wire References

CC does NOT auto-load references/ files. The model never sees them unless SKILL.md points to them explicitly.

1. Add inline "read this when" links at the point of use.
2. Add a ## References table at the end for 3+ files.
3. Verify every file in references/, scripts/, assets/ is reachable — orphan files are invisible.

6. Frontmatter Validation

Read ${CLAUDE_SKILL_DIR}/../prompt-writer/references/frontmatter-schemas.md for the complete skill frontmatter schema. Key effective fields: name, description, when_to_use, user-invocable, disable-model-invocation, effort, context, agent, argument-hint, arguments, paths, hooks. Do NOT declare model, allowed-tools, disallowedTools, or tools — CC ignores them for plugin components.

7. Review

Present the draft. Run the quality checklist below before delivering.

Progressive Disclosure Architecture

Design every skill with three loading levels.

Level 1 — Metadata (~100 tokens)

• Only name + description injected into <system-reminder> listing
• Hard cap: 250 characters per entry — front-load key use case
• This is ALL the model sees until it invokes the skill

Level 2 — SKILL.md body (≤5k tokens)

• Full body injected on invocation via Skill tool or /name
• On context compaction: truncated to 5,000 tokens with ellipsis
• Target: <500 lines / <3k words. Challenge every token — the model already knows general concepts

Level 3 — Bundled resources (unlimited)

• references/ — detailed docs, API refs, schemas
• scripts/ — executable helpers (stdout consumed, source NOT loaded)
• assets/ — HTML templates, static files
• NOT auto-loaded — model must explicitly Read them
• One level deep rule: SKILL.md → reference.md works. Deeper chains break

Token Efficiency

• SKILL.md body under 500 lines / 3k words
• Large references (>100 lines) need a TOC at the top
• Scripts in scripts/ — executed, not loaded (only stdout consumes tokens)
• Avoid time-sensitive information in the body
• Move API refs, schemas, and boilerplate to references/

Quality Checklist

Run before presenting any skill to the user.

Frontmatter

1. Description ≤250 chars, front-loaded with key use case?
2. when_to_use present for auto-triggered skills?
3. user-invocable / disable-model-invocation set correctly?
4. No ineffective frontmatter declared (model, allowed-tools, disallowedTools, tools)?
5. paths field used if file-type specific? context: fork if parallelizable?

Content 6. Every sentence is an action? ("Consider" → "Use", "Extract", "Run") 7. No unnecessary concept explanations? (the model already knows them) 8. No vague words? ("good", "proper" → concrete rule) 9. No repetition across sections? 10. CRITICAL/IMPORTANT inflation? Maximum 1-2 per file

Progressive Disclosure 11. SKILL.md body <500 lines? References >100 lines have TOC? 12. Every reference explicitly linked with "read this when" guidance? 13. All files in references/ reachable from SKILL.md? 14. One level deep only — no chained reference files?

Style 15. Imperative mood throughout? No passive voice? 16. Nesting depth ≤1? Sentences under 25 words? 17. No hardcoded paths or usernames?

References