draft-skill

An AI agent skill that drafts a new skill — a SKILL.md for Claude Code, Codex, Cursor or Qoder — from a plain idea: no prompt list, no zip, just "I want an agent that does X". It confirms the purpose before writing a file, checks the idea against the skills already installed so two of them don't end up claiming the same request, decides what belongs in scripts/ versus references/ versus assets/, writes a description that actually triggers, validates the result with a script, and hands it off for tuning and packaging.

The two things it is strictest about are the two that fail silently in production: a description that promises a step the body never performs, and a name that already belongs to a better-established project. Both have a check written down rather than a feeling.

Siblings: prompts-to-skill when you already have a list of prompts, tune-skill to fix a skill that exists. Anthropic's skill-creator covers the same ground with evals; this one is plain markdown plus a validator. More at skills.caffeinity.com/draft-skill.

Install

npx skills add arm-pz/draft-skill --skill draft-skill

Or copy the draft-skill/ folder into ~/.qoder/skills/ (Qoder) or ~/.claude/skills/ (Claude Code), then start a new session.

It calls check_skill.py from tune-skill for validation. Install that first, or the validate step has to be done by hand.

Ask it things like

• "Make me a skill that reviews a client's SEO report."
• "I keep pasting the same instructions — turn that capability into a skill."
• "Scaffold a skill for talking to our internal API."
• "There's an empty folder in ~/.qoder/skills/ — fill it properly."

What it does before writing anything

1. Collects three or four requests the skill should catch and one it should refuse, plus the output shape.
2. Reads the description of every installed sibling the examples touch, from the file.
3. Sets the boundary in both directions — what this owns, where it stops, who takes over.
4. Plans resources, then names the skill after checking GitHub for collisions.

Step 4's collision reading is documented in references/structure.md, including the trap that github.com/topics/<name> renders a page for names that don't exist, so only a topic: search count is evidence.

Layout

draft-skill/
├── SKILL.md                    # the 8-step authoring workflow
├── LICENSE
└── references/
    └── structure.md            # collision reading + the 7 checks the script can't make

Boundaries

• Handles: a new skill, from an idea.
• Does not handle: a batch of existing prompts (prompts-to-skill), tightening a skill that already exists (tune-skill), releasing a finished one (prompts-to-skill-release), or making the repo findable (findable).
• Serves: the create step of the pipeline draft-skill → tune-skill → findable.

Provenance and status

Written from scratch on 2026-10-19 for this skill suite. Its rules are the conventions already enforced by the sibling skills in this repo family — the description shape and the test-case format from tune-skill, the confirmation gates from prompts-to-skill — plus the deterministic checks check_skill.py runs. A skill-creator is the conventional first skill in the Agent Skills ecosystem, and this one is addressed to that idea; it contains no text from Qoder's built-in skill-creator, which ships without a license permitting redistribution, so nothing from it was copied, reworded or restructured here.

SKILL.md passes check_skill.py: 0 FAIL, 0 WARN, 783 description characters, 69 body lines. The skill's own trigger test cases have not been run in a live session, and no evals/ file is shipped.

License

MIT. See LICENSE.