New Skill Creator

Interactive wizard to scaffold a new Claude Code agent skill from scratch.

Workflow

1. Gather requirements — capability type, scope, tool access via AskUserQuestion
2. Name & describe — kebab-case name + trigger-rich description
3. Generate content — pick template, customize for project context
4. Create files — write SKILL.md (+ reference files for advanced skills)
5. Confirm — show file tree and example activation phrases

Requirements Gathering

• Ask three targeted questions (capability, scope, tools) before writing any files
• Never assume capability type or tool access — always confirm
• Use AskUserQuestion tool with 2-4 options per question
• Read reference for exact question structure

References: requirements-wizard

Naming & Description

• kebab-case, max 64 chars, describe what it does not how
• Description must include: functionality + trigger scenarios + keywords
• Template: [Primary function]. Use when [scenarios]. Keywords: [terms].
• Max 1024 chars for description; include 5-10 trigger keywords

References: naming-and-descriptions

Skill Content

• Choose the template matching the capability (code analysis, test gen, docs, data)
• Customize for project tech stack from deno.json, CLAUDE.md, existing patterns
• For advanced scope: use reference-file pattern — lean SKILL.md + references/ dir
• Skills with >150 lines of content should split into reference files

References: skill-templates

File Structure & Tools

• Simple: .claude/skills/<name>/SKILL.md only
• Advanced: add references/, scripts/, or templates/ subdirectories
• Add allowed-tools: frontmatter if user chose restricted tool access
• Common combos: read-only = Read, Grep, Glob; data = Read, Write, Bash(jq:*)

References: file-structure-and-tools

Confirmation

• Show created file tree with ls -R .claude/skills/<name>/
• List example phrases that would activate the skill
• Explain why the description will trigger reliably
• Suggest a test request the user can try immediately

References: best-practices

Guardrails

• Do not create files until all three requirement questions are answered
• Do not hard-code machine paths or secrets in generated skills
• Generated skills with non-trivial content (>150 lines) should themselves use the reference-file pattern
• Always read existing CLAUDE.md and project context before generating skill content