Skill Creator

Create a new Claude Code skill following standard conventions.

Process

1. Ask the user:

   • Name: in category-name format (e.g. coach-call, product-prd, ops-deploy)
   • Purpose: what should this skill do? (1-2 sentences is enough)
   • Allowed tools: which tools can the skill use without asking permission? (default: Read, Write)

2. Create the skill folder and file:

   • Path: <skills folder>/{category-name}/SKILL.md (the skills directory, typically symlinked to ~/.claude/skills/)
   • The dash separator is sync-safe: tools that block : in folder names (some note editors) handle a dash fine
   • Run ls on the skills folder first to check existing categories

3. Write the SKILL.md following this template:

---
description: [One clear sentence. Start with a verb. Explain WHEN to trigger this skill.]
allowed-tools: [Comma-separated list of tools]
---

# [Skill title]

[Core instructions for Claude when this skill is invoked.]

## [Sections as needed]

[Keep instructions precise and actionable. No filler.]

Conventions

• Language: all skill files are written in English (instructions, frontmatter, comments)
• Folder naming: category-name using a dash separator. Dash (not colon) so the folder syncs safely across editors and devices, including tools that block : in folder names
• Description field: starts with a verb, explains when to trigger, one sentence. A vague description ("Helps with documents") gives the agent no way to distinguish this skill from others — name the concrete triggers
• Tone of instructions: direct, imperative, no fluff
• Location: your skills folder (typically symlinked to ~/.claude/skills/)
• After creation: the skill is immediately available as a slash command /category-name in Claude Code

Progressive disclosure

Keep SKILL.md concise. As a rule of thumb, if it grows past ~100 lines, split the detail into companion files loaded on demand:

• REFERENCE.md, EXAMPLES.md, FORMAT.md — pulled in only when the relevant step runs
• scripts/ — for deterministic logic. A bundled script saves tokens and is more reliable than code regenerated each run

Keep references one level deep (SKILL.md points to companion files, companion files do not chain further).

Categories already in use

Check existing skills before creating:

ls ~/.claude/skills/

Common categories:

• writing- — text writing, editing, coaching
• product- — product management, PRDs, analysis
• coach- — structured thinking, calls, feedback
• ops- — infrastructure, server, deployment
• veille- — monitoring, digests, triage
• mktg- — marketing copy, CRO, expert panels

Review checklist

Before finishing, verify:

• Description starts with a verb and names concrete triggers
• Folder uses the category-name dash convention
• No time-sensitive info baked in (dates, "current" versions)
• Terminology is consistent throughout the file
• SKILL.md stays concise; detail externalised if past ~100 lines
• References are one level deep
• Frontmatter has both description and allowed-tools

What NOT to do

• Never use a language other than English in the SKILL.md content
• Never skip the frontmatter (description + allowed-tools are required)
• Never bake time-sensitive info into a skill (it goes stale)
• Never create a skill without confirming name and purpose with the user first