skill-builder
Use when turning a repeated procedure or SOP into a skill, or deciding whether a step should stay a prompt, become a skill, a loop, or a scheduled routine.
SKILL.mdpython_skills/llm_instructions/chat_skills/skill-builder/SKILL.md
GitHubSkill builder
From the user's playbooks "Turn an SOP into a Claude Skill" and "Prompt to skill to loop to routine", plus the course units on skills (L13-22 contract, L13-24 progressive disclosure, L13-25 routing, L13-27 evals and packaging).
1. Pick the level (the automation ladder)
- Prompt: run a new step by hand until it reliably works. Do not make a skill of something that has not worked twice.
- Skill: steps that keep recurring (video intake, design intake, a curation pass) get saved so one request repeats them.
- Loop or routine: only steps that are safe unattended (a weekly check for entries marked unchecked) get scheduled, and they must report failures somewhere the user looks.
- Never automate judgment: deciding what goes into 411, accepting claims, and the final review stay with the user.
2. Write the SOP first
Plain steps in order: trigger, inputs, each step, the check that proves it worked, the output, what it must not do. If the SOP cannot be written, the skill is not ready.
3. Convert to SKILL.md
- Front matter:
name(kebab-case) and a one-sentencedescriptionthat says WHEN to use it. Keep it short, because it is what loads first (progressive disclosure); detail goes in the body. - Body: the steps, rules, output format, and which other skills it hands off to.
- Python does the mechanical parts (counting, checking, fetching, frames); the skill names the script and its expected output.
4. Keep the two kinds distinct
- Claude chat skills are saved to the account through the skill proposal card. They must not assume a shell on the laptop or access to sites chat cannot reach (X).
- Claude Code skills live in
.claude/skills/<name>/SKILL.mdon the laptop, with scripts beside them or in the repo, and a copy in pyburticapapython_skills/llm_instructions/.
5. Test, fix, repeat
Run it end to end on a real case, fix each failure in the SKILL.md (not in chat), run again. A skill is done when a fresh session follows it without extra explanation.
Rules
- Read an existing skill in full before changing it; a saved change replaces the whole file.
- One skill per job; link to other skills instead of copying their steps.
- Labels on any facts inside a skill: verified here, source claim, our inference.