chat-to-skill-publisher

Chat to Skill Publisher

Trigger

• Use when the user wants to turn a conversation into a reusable skill package.
• Use when the current in-agent chat already contains a stable workflow that should be saved as a skill.
• Use when the user provides optional pasted transcript text from Claude or another agent to supplement the current chat.
• Do not use this skill for small edits to an existing skill unless the main job is still deriving a new reusable skill from chat history.
• Do not use this skill when the conversation has no repeatable workflow yet; refine the workflow first.

Inputs

• project_dir: current project root where the draft should be staged under .temp/<skill-id>/
• skill_brief: short instruction describing what kind of skill should be extracted
• transcript_text: optional pasted transcript text from Claude or another agent
• skill_id_hint: optional preferred skill id
• marketplace_repo_url: optional target marketplace repository URL, defaulting to https://github.com/zhouyou-gu/skill-marketplace-template
• marketplace_local_path: optional filesystem path to a local clone of the marketplace repository; when set, it is used as the source for parent inference and for publish
• publish: when true, publish after drafting
• dry_run: when true, return the draft and publish plan without writing

Workflow

1. Read the current chat history and any transcript_text, then isolate the reusable workflow: triggers, inputs, outputs, failure rules, and concise normalized examples.
2. Interact with the user when the workflow is ambiguous or when too much of the transcript is one-off context rather than reusable instruction.
3. Resolve the skill id from skill_id_hint or from the workflow itself. Keep it lowercase kebab-case and consistent across folder name, SKILL.md, and skill.yaml.
4. Draft a full skill package in project_dir/.temp/<skill-id>/: SKILL.md, skill.yaml, tool.json, README.md, and examples/.
5. If project_dir is a git repository and .temp/ is not ignored, prefer adding a local-only ignore via .git/info/exclude instead of editing tracked ignore files. If that is not possible, continue with a warning.
6. Genericize the draft (mandatory, see "Genericize Before Publish" below). The chat history the draft was derived from is shaped by one specific project; the published skill is not. Review every file in the draft and rewrite project ties so the skill reads as a reusable workflow, not a replay of the originating session. This step runs before parent inference because a non-generic draft mis-matches existing parents.
7. Infer parent skills from the target marketplace (see "Parent Inference" below) and update the draft in place: record the confirmed ids on inferred_parents, write them into the draft's skill.yaml under parents, and add a matching ## Parents section to the draft's README.md with one line per parent explaining the relationship.
8. Keep the draft aligned to skill-marketplace-template conventions: matching ids, allowed category, required metadata, valid tool contract, non-self-referential parents, and concise reusable examples.
9. If publish is false, stop after drafting and return the draft summary.
10. If publish is true, resolve the marketplace target from marketplace_local_path or marketplace_repo_url by preferring a matching local clone. If a local clone is available, publish into skills/<skill-id>/ there and run the validation and registry build checks. If no local clone is available, keep the draft and return a publish-ready summary with warnings.