Skill Creator

Interactive 8-step guide for creating, reviewing, and fixing custom skills. Structured around Anthropic’s Complete Guide to Building Skills for Claude (Jan 2026) with the three-level Progressive Disclosure model baked into every step.

Metadata

Field	Value
Type	command
Invoked by	`/skill-creator`, or asking to create a skill, build a skill, make a new slash command skill, add a coding standards skill, review an existing skill, update a skill, or fix a skill that doesn’t trigger

Included files

Path	Description
`assets/`	Templates and reusable content
`references/`	Reference documentation for Claude

Usage

/skill-creator

Or natural language:

Help me build a new skill for deploying to our staging environment

I have a skill that never triggers — can you review it?

The 8-step flow

Step	Focus
1	Gather 2–3 concrete use cases
2	Pick skill type (command vs. context)
3	Write name + description with trigger phrases
4	Frame as problem-first vs. tool-first
4b	Pick a pattern (sequential / multi-MCP / iterative / context-aware / domain-specific)
5	Decide folder structure (just `SKILL.md`, or add `references/` / `assets/` / `scripts/`)
6	Write the `SKILL.md` body
7	Test triggering / functional / performance
8	Save, activate via `./install.sh --refresh-custom`, verify

Progressive Disclosure

The skill uses the three-level model from Anthropic’s spec:

Level 1 — YAML frontmatter (always loaded): name + description — Claude’s trigger matcher reads this on every request. Under 1024 characters.
Level 2 — SKILL.md body (loaded when relevant): decision flow, critical rules. Under ~5000 words.
Level 3 — Linked files (loaded on demand): references/ for docs, assets/ for templates, scripts/ for deterministic code.

Each reference file in the bundled eight covers one level-3 topic in depth — load only what the current task needs.

Reference files

Reference	Covers
`01-progressive-disclosure.md`	The 3-level system; when to use references / assets / scripts
`02-frontmatter.md`	All YAML fields, security rules, reserved names, harness-vs-spec distinction
`03-writing-descriptions.md`	Trigger phrases, what+when structure, good/bad examples
`04-writing-instructions.md`	Specificity, structure, error handling, when to reach for scripts
`05-patterns.md`	The 5 common patterns with full example structures + Skills+MCP story
`06-testing.md`	Triggering / functional / performance testing methods, ≥90% benchmark
`07-troubleshooting.md`	Upload errors, trigger diagnosis, MCP issues, large-context issues
`08-checklist.md`	Pre-upload and post-upload checklists
`09-distribution.md`	Sharing skills via Claude.ai, Messages API, and GitHub

Worked examples

Three asset templates you can copy and modify:

skill-template.md — minimal SKILL.md scaffold with placeholders
command-skill-example.md — a complete deploy-staging command skill with real kubectl / docker commands, error handling, rollback
context-skill-example.md — a complete Python coding-standards context skill covering naming, FastAPI conventions, SQLAlchemy, pytest

Skill types

The claude-code-setup harness recognises two invocation styles on top of the Anthropic Skills spec:

Type	Frontmatter	How it loads
Command	`type: command`	User invokes with `/skill-name`
Context	`type: context` + `applies_to: [...]`	Harness auto-loads when the project’s Tech Stack matches

Both produce folders that are portable — the same folder works on Claude.ai, Claude Code, and the Messages API. The type / applies_to / file_extensions fields are harness-specific and are ignored by upstream Anthropic surfaces.

Output location

Skills created through the flow are saved to:

~/.claude/custom/skills/<skill-name>/

This location is separate from the base installation (~/.claude/skills/) and survives claude-code-setup upgrades. Activate a newly created skill with:

./install.sh --refresh-custom

or via /claude-code-setup → Upgrade custom.

Creating Skills — conceptual guide to skill architecture
Customizing — share custom modules with your team

Slidev Presentations Gradle