# Skill Create

> Create or update repository-local Codex skills under .codex/skills. Use when the user asks to create a skill, update a SKILL.md, change skill metadata, add skill references/rules/templates, validate skill behavior, or fix a skill's required layout.

- Skill: `yenhao-huang/skill-create` (Agent Skill, multi-file: 9 files)
- Install (CLI): `npx skillmds@latest add yenhao-huang/skill-create`
- Raw SKILL.md: https://api.skillmd.com/api/skills/yenhao-huang/skill-create/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: yenhao-huang (https://skillmd.com/u/yenhao-huang)
- Updated: 2026-09-21
- Page: https://skillmd.com/skills/yenhao-huang/skill-create

---


# Create Skill

Use this skill to create and maintain repo-local Codex skills. In this
repository, a valid skill has a stricter layout than the generic system skill
initializer creates by default.

## Workflow

1. Read the user's request and identify the exact skill name, purpose, and
   target directory.
2. Read `STATE.md`; for a new run, reset it from
   `references/template/STATE.template.md`.
3. Read the required references:
   - `references/rules/categories.md` before choosing a target directory or
     moving an existing skill.
   - `references/rules/filetree.md` before adding, moving, or removing files in
     any skill directory.
   - `references/rules/workflow.md` before creating or changing a skill's
     workflow, installation steps, rollback behavior, or validation plan.
   - `references/rules/env.md` before installing packages, starting services,
     choosing frameworks, or assuming runtime details.
   - `references/rules/state-rules.md` before changing any `STATE.md`.
4. Mark the current step `in_progress` in `STATE.md` before editing.
5. Create or update the skill under the category selected from
   `references/rules/categories.md`, using the local required layout from
   `references/rules/filetree.md` and the workflow contract from
   `references/rules/workflow.md`.
6. Require an explicit `## Workflow` section. When a skill is intended to be
   reversible, pair every step that installs or enables component A with
   instructions that uninstall, disable, or otherwise roll back that same A.
7. Validate with generic skill validation when available:
   `/home/howard/.codex/skills/.system/skill-creator/scripts/quick_validate.py`
   Then inspect the skill directory against `references/rules/filetree.md` and
   `references/rules/categories.md`. For each reversible installation path,
   also run and record the full `install A -> rollback A -> install A` sequence
   required by `references/rules/workflow.md`.
8. Mark the workflow step `completed`, `blocked`, or `skipped` in `STATE.md`
   with concrete evidence.

## Conflict Rule

If a generic system initializer or external template conflicts with
`references/rules/filetree.md`, the repo-local filetree rule wins.

## Skill Layout Contract

Every newly created or substantially updated repo-local skill must include:

- `SKILL.md`
- `STATE.md`
- `references/rules/filetree.md`
- `references/rules/env.md`
- `references/rules/state-rules.md`
- `references/template/STATE.template.md`

`SKILL.md` must contain a `## Workflow` section with executable ordered steps.

Additional domain references are allowed under `references/`.
`references/scripts/` is the place for executable code needed by the skill.
Do not add `agents/` or top-level `scripts/` as part of the default
repo-local skill layout.

## Reference Rules

- Keep `SKILL.md` focused on trigger conditions, core workflow, and short
  guardrails.
- Put detailed domain behavior in `references/`.
- Put skill executable code in `references/scripts/`.
- Put reusable operating rules in `references/rules/`.
- Put state templates in `references/template/`.
- Require runtime evidence for reversible installation workflows; prose-only
  inspection does not satisfy the `install -> rollback -> install` check.
- Keep skill directories under the approved category taxonomy.
- Load only the reference files needed for the current request.
- Do not add README, changelog, installation guide, or quick-reference files
  unless the user explicitly asks for user-facing documentation.

## Output

Final responses should include:

- What skill files changed.
- Which validation commands passed.
- Whether the target skill now satisfies the local required layout.
- Any blocker or remaining risk.