skill-factory
Builds a skill the way a careful maintainer would, then holds the release until the user approves it. Works in any agent; each step names the best tool and a fallback that always works.
Input: a skill idea, or a drop handed over by loot with its evidence (query,
search results, proof, edge). Target folder: where the user says, else
../<slug> outside the current repo's history.
1. Spec
Write SPEC.md: goal, inputs and outputs, and an acceptance checklist where
every line can be tested. Under a page. If loot handed the item over, reuse its
evidence instead of searching again. Ask the user only when two readings of the
idea would produce different skills.
2. Build, tests first
- If the agent has a skill-authoring helper (in Claude Code:
anthropic-skills:skill-creator), use it for the skill and its evals. - Otherwise write
SKILL.mdyourself: frontmatter withname(kebab-case) and adescriptionthat says what it does and when to trigger, then the steps. - Scripts get tests before code (
node:testor pytest). Implement until they pass. Write files with your file-editing tool, not shell heredocs. - Run every test and eval. Keep the exact command and its output for step 4.
3. Independent review
The builder doesn't grade its own work.
- Best: start a subagent that has not seen this conversation. Give it only
the target folder and
references/release-checklist.mdwith this brief: "Review the skill in against the checklist. Answer PASS or FAIL per item with file:line evidence. When unsure, FAIL. Flag any claim the tests or evals don't prove." - Fallback: no subagents: for each checklist item, write the strongest case that it FAILS before deciding.
Fix every FAIL, rerun tests, review again. Stop when all items pass, or when an item is blocked on the user (say which and why).
4. Release gate
Scan for secrets: node <loot skill folder>/scripts/scan.mjs <folder>, plus
gitleaks dir <folder> if installed. Then show the user a manifest and stop
for a clear yes:
- repo name, public or private
- version and changelog line
- the files to publish
- test and eval results (command + pass count), the review verdict
- anything skipped
A yes covers exactly that manifest. Then publish from a fresh git init:
commit release vX.Y.Z, tag, gh repo create and push, and install the skill
where the user's agent loads skills. No gh or not signed in: print the
commands. Never ask for, type or store credentials.
5. Report
One short message in plain language: links, test results, review result, and what's left for the user.