skill-create-workflow

Turn a reusable idea, postmortem, or prompt into a new skill — three gates (content-fit, naming, prompt review) run before anything hits disk.

When to use it

Fresh idea worth crystallizing:

Something in your head or in a doc — a postmortem, a checklist, a team rule, an agent's system prompt — keeps getting reused, and you want to turn it into a skill so a one-liner triggers it next time.

Follow-up "yes":

The last round of chat already asked "is this worth a skill?" and you reply with a short "yeah / do it / go for it". The flow picks up the discussed material as the source without asking for a path again.

Existing material at a path:

You hand over a docs page / postmortem / internal wiki entry by path. The flow wraps it into a fresh skill directory (name, frontmatter, triggers, workflow, examples, failure paths).

Create and publish in one shot:

Once the new skill lands you don't want to branch / commit / open a PR by hand — you want the flow to push it to the company skill library on the way out.

Out of scope:

Review or improve an existing skill → skill-improve-workflow; only pick a display name → skill-name-generation; only pick a slug and boundaries → skill-domain-framing; only judge whether the content is worth a skill without building it → skill-content-fit; install / list / upgrade / uninstall an installed skill → tranfu-router.

What it produces / you'll see

Four gates run before anything hits disk — fail any gate and the flow stops at the gate instead of silently inventing a bad skill. This is the most counterintuitive part.

• Content gate: skill-content-fit decides whether the material is worth becoming a skill at all; a 打回 result stops the flow before naming
• Container & boundary: skill-domain-framing picks the slug, include / exclude scope, and placement; if the top-two candidates are within 2 points the flow stops and asks you to choose
• Skill files on disk: skill-creator (Codex system one, or Claude Code's native authoring) writes only under {owned_skill_directory} — SKILL.md plus any agents/ directory
• Prompt review loop: prompt-review audits every file that carries a prompt for engineering quality; up to 5 rounds; if it still fails the flow stops and asks you
• Publish leg: tranfu-publish uploads to the company skill library (git commit / push / PR); non-blocking — a publish failure never rolls back the skill
• Never does: touch unrelated skills, project code, or hand-written files; skip any gate; merge your PR

Prerequisites & boundaries

Prerequisites:

Requires skill-content-fit, skill-domain-framing, skill-creator, and prompt-review to be available in the current runtime; missing any one → the flow reports the exact missing command and stops, never installs or scrapes an unofficial template. On Codex the "stop and ask" path uses request_user_input when Plan mode exposes it, otherwise falls back to a numbered plain-text prompt. The publish leg additionally needs tranfu-publish plus a working git environment.

Adjacent skills:

Task	Route to
Health check / improve an existing skill	skill-improve-workflow
Pick display names (English + Chinese)	skill-name-generation
Pick a slug and its boundaries	skill-domain-framing
Only judge whether content is worth a skill	skill-content-fit
Publish an already-built skill	tranfu-publish
Install / list / upgrade / uninstall an installed skill	tranfu-router

Out of scope:

• Installing / listing / upgrading / uninstalling skills
• Creating a plugin
• Editing ordinary project code or writing non-skill documentation
• Managing scheduled tasks

Subtle edges:

• Previous turn discussed "is this worth a skill?" and you reply "yeah / go for it" → triggers this skill; the discussed content is the source, no path required
• Pronoun reference ("it / this / that thing above") plus a create verb → triggers this skill; resolve context instead of refusing on "no path given"
• Add a specific feature to an existing skill → this skill in update mode; full quality review of an existing skill → skill-improve-workflow