create-skill

Create a New Skill

Where to place it

Choose the location based on scope — in order of priority:

After writing the file:

1. Add a ## Skills section (or entry) to the corresponding CONTEXT.md — root .ai/CONTEXT.md for cross-cutting skills, or .ai/Component/{Name}/CONTEXT.md / .ai/Domain/{Name}/CONTEXT.md for scoped skills. This is the agnostic discovery mechanism for all non-Claude tools.
2. Create a symlink in .claude/skills/ pointing to the skill directory (not the file). The path must be relative. This enables Claude Code auto-discovery.

   cd .claude/skills && ln -s ../../<skill-dir-path-from-repo-root> <skill-name>
   

---

CONTEXT.md vs SKILL.md — no duplication rule

CONTEXT.md owns conventions (rules, patterns, constraints — the "why" and "what"). SKILL.md owns procedures (steps, code templates, checklists — the "how"). Content must live in exactly one place.

When creating or modifying a skill

For every rule or convention you're about to write in a skill, apply this decision:

1. Read the parent CONTEXT.md first — before writing any content in a skill, read the component (or domain) CONTEXT.md to know what's already documented there
2. Ask: "Does this apply to all skills in this component?" — if yes, it belongs in CONTEXT.md, not the skill. Example: "all handlers use #[AsCommandHandler]" → CONTEXT.md
3. Ask: "Is this specific to this one task?" — if yes, it stays in the skill. Example: "the edit handler checks null before each field" → skill
4. If a convention is missing from CONTEXT.md and should be there — add it to CONTEXT.md first, then reference it from the skill. Never write it only in the skill
5. Reference, don't restate — when a skill needs to remind the reader of a convention, write: See [Component/CONTEXT.md](../../CONTEXT.md#section) for X convention. Do not copy the rule text into the skill

When reviewing an existing skill

Check the skill's ## Rules section and any inline convention statements:

1. For each rule in the skill, check if the same rule exists in the parent CONTEXT.md
2. If it does → delete it from the skill and replace with a reference
3. If it doesn't but should → move it to CONTEXT.md and replace with a reference
4. If it's genuinely task-specific → keep it in the skill

What NOT to put in CONTEXT.md

Not everything belongs in CONTEXT.md either. Skip:

• Class inventories, file listings — anything grep or glob can answer
• Code templates and step-by-step procedures — those are the skill's job
• Content already in the root .ai/CONTEXT.md (project-wide coding standards, testing framework)

Cross-references and cascade risk

The ## Related section in CONTEXT.md files links to other contexts. Use it sparingly — every link is a potential cascade where an AI agent follows A → B → C and ends up loading all contexts, which defeats the purpose of splitting them.

Only link when the relationship is non-obvious (architectural surprise, coexistence gotcha). Do not link for obvious usage relationships ("Controller uses CQRS") or just to mention a class name that's greppable. When in doubt, omit the link.

---

SKILL.md format reference

Frontmatter (YAML between --- markers)

Project-specific frontmatter (custom metadata)

This project uses four custom frontmatter fields on top of the standard ones: needs, produces, conditional, subagent. They are not interpreted by Claude Code but document the skill's place in larger workflows and serve as machine-readable hints for orchestrators.

See STRUCTURE.md → SKILL.md project conventions for the full definition of these fields, the top-down dependency rule, and the standalone rule. Do not duplicate that documentation here.

Arguments: If a skill requires arguments that the user must provide, describe them in the description field. At runtime, if a required argument is missing, use AskUserQuestion to prompt the user.

Body

Plain markdown. No required sections — write whatever instructions I need to follow.

Useful substitutions in body:

• $ARGUMENTS — full argument string passed after the skill name
• $0, $1, … — individual arguments by position
• !`command` — runs a shell command before I see the content; output is inlined

Invocation behaviour

---

Minimal template

---
name: my-skill
description: >
  One sentence what it does. Trigger phrases: "do X", "create Y", "add Z".
allowed-tools: Read, Grep
---

# My Skill

## Purpose
What this skill accomplishes.

## Steps
1. …
2. …

## Output
Where to write the result.

---

Upstream reference

The canonical Anthropic skill creator (may contain updates not yet reflected here): https://github.com/anthropics/skills/blob/main/skills/skill-creator/SKILL.md

If that file has evolved, reconcile any new fields or patterns with the project-specific placement rules above before applying them.

---

Checklist

• Directory and SKILL.md created at the correct scoped path (component, domain, or cross-cutting)
• description front-loads the use case and lists trigger phrases
• description does not restate "Read Component/X/CONTEXT.md…" — that line belongs in the body only
• If the skill needs the parent CONTEXT.md, the body opens with a single Read @.ai/Component/{Component}/CONTEXT.md for ... reference
• Custom frontmatter fields (needs, produces, conditional, subagent) follow STRUCTURE.md conventions — top-down dependencies, standalone-usable
• If the skill is read-heavy and produces a structured artifact, consider declaring subagent: recommended or subagent: optional (see STRUCTURE.md). Absence is the default — leave the field off if the skill is small, interactive, or cross-references siblings
• Corresponding CONTEXT.md updated with a ## Skills entry (agnostic discovery)
• Symlink created in .claude/skills/ pointing to the skill directory (Claude Code discovery)