| name | skill |
| description | Turns a repeated workflow into one narrow, reusable, validated agent skill with reliable triggers and progressive disclosure. Use when the user asks to create, revise, split, review, or test a SKILL.md workflow for Claude, Codex, or another Agent Skills-compatible surface; not when the request is to distribute a bundle as a plugin. |
Skill
Create a skill around one repeatable user intent and one output contract.
References
Resolve the target product, then read its branch:
- Claude:
references/claude-code-estensioni.md.
- Codex:
references/codex-skills.md.
- Read
references/ai-fluency-4d-corso.md when delegation, discernment, or diligence rules shape the workflow.
Procedure
- Confirm the workflow is repeatable and belongs in a skill rather than always-on instructions, a subagent, hook, plugin, or MCP integration.
- Define at least three representative triggers and clear non-triggers.
- Specify inputs, one primary output, failure behavior, side-effect policy, and success checks.
- Keep essential procedure in
SKILL.md; place detailed knowledge in directly linked references/; use scripts only for deterministic repeated operations.
- Match freedom to risk: guidance for flexible work, templates for repeatable work, scripts for fragile deterministic work.
- Follow the selected product's current naming and packaging rules.
- Validate structure and forward-test triggers, non-triggers, normal cases, boundaries, and failures.
Output contract
Return:
- Decision — why a skill is the right mechanism.
- Contract — triggers, non-triggers, inputs, output, failure behavior, and authority.
- File tree — only required files.
- Implementation — complete files when requested.
- Validation — exact checks and forward-test prompts.
Boundaries
- Do not create an “everything” skill.
- Do not package or publish a bundle of existing workflows; route distribution work to
plugin.
- Do not duplicate detailed material in both
SKILL.md and references.
- Do not add repository documentation inside a skill folder.
- Do not claim success before validation and trigger testing.
- Answer in the user's language.
