Créer des assets AgentKit (skills, commandes, agents)

---

description: Scaffold a new skill, command, agent, or Cursor rule with correct frontmatter and location. argument-hint: <kind> <name> [description] allowed-tools: Bash, Read, Write, Edit, Glob

Scaffold a new AgentKit asset. Parse $ARGUMENTS as: <kind> <name> [description].

• <kind> = skill | command | agent | rule | guideline
• <name> = kebab-case (my-skill, backend-engineer, go-service)
• [description] = optional one-liner; ask for it if missing

Where to create it

First determine context: are we inside an AgentKit checkout (has skills/, commands/, agents/ at the root), or inside a consuming project with .claude/ installed?

| Context | Path | | --- | --- | | Consuming project (.claude/ present) | .claude/skills/<name>/SKILL.md etc. | | AgentKit library root | skills/<name>/SKILL.md etc. |

Full dispatch table (consuming project):

| kind | Creates | | --- | --- | | skill | .claude/skills/<name>/SKILL.md | | command | .claude/commands/<name>.md | | agent | .claude/agents/<name>.md | | rule | .cursor/rules/<NN>-<name>.mdc (ask for NN prefix) | | guideline | shared/guidelines/<name>.md (AgentKit only) |

Before writing

1. Check for collisions. If the target path exists, stop — don't overwrite.
2. Validate the description. Must be trigger-shaped: "Use when the user asks to…". Reject summary-shaped descriptions like "Handles X-related tasks."
3. For rules, ask for the globs array — don't guess the file pattern.

Use the template

Find the right template from templates/ (or from where AgentKit is installed):

• Skill → SKILL.template.md
• Command → command.template.md
• Agent → agent.template.md
• Rule → rule.template.mdc

Fill every <placeholder>. Mark genuinely optional sections with <!-- TODO -->.

After creating

1. Print what was created and what still needs filling in (the <...> fields).
2. Remind the user:
   • Description must be trigger-shaped or discovery won't work.
   • No project-specific names, credentials, or internal URLs in assets meant to be reused.
   • Reference shared/guidelines/ for security, multi-cloud, or document rules rather than duplicating them inline.
   • Keep the body under ~120 lines — it's loaded into context on every use.