harness-creator

This repo builds harness-creator, a meta-skill that designs, generates, and validates Claude Code harnesses (CLAUDE.md, rules, skills, hooks, permissions, agents, workflows) for other projects through a structured interview.

Where things live

• The skill itself: .claude/skills/harness-creator/ — SKILL.md is the orchestrator, references/ holds per-component authoring guides, scripts/ holds the Python CLIs it calls.
• The record of what was decided and why: CHANGELOG.md's [Unreleased] section plus the commit and pull-request bodies behind it. There is no separate spec file — the generation plans under docs/plan/ were removed along with .claude/harness-spec.md, for the reason the changelog gives: a record nothing keeps honest can be silently wrong, which is worse than none for a skill whose value is that its gotchas are true. Git is often incomplete but never wrong about what changed. If you change a decision, say in the commit body which decision it supersedes, since history cannot be edited in place.
• Self-test fixtures for the scripts: tests/fixtures/ at the repo root, not inside the skill — anything under the skill directory ships to plugin users, so dev-only fixtures stay outside it.
• Local reference snapshots the plan was researched against: .tmp/docs_claude/ and .tmp/skill-creator/ — gitignored, not part of this repo's own history.

Working on this repo

• This repo is dogfooding its own output shape: .claude/skills/harness-creator/ is simultaneously (a) a project skill for this repo and (b) the plugin's skill component (.claude-plugin/plugin.json's skills field points here). Don't duplicate the skill elsewhere.
• Daily development loop is the symlink, not the plugin install: ~/.claude/skills/harness-creator → this repo's skill directory. Edits are reflected immediately in a fresh session. Only install the plugin via the local marketplace when smoke-testing the distribution path itself, and uninstall it afterward — running both at once double-registers the skill under two different names (bare harness-creator vs. harness-creator:harness-creator).
• No mid-sentence hard line-wraps in any markdown written here (SKILL.md, references, this file, generated harness output). Line breaks only at sentence/list/paragraph boundaries — hard wraps break Edit-tool string matching and pollute diffs, and renderers soft-wrap anyway.
• Every claim the skill makes about Claude Code mechanics (hook exit codes, matcher syntax, CLAUDE.md loading semantics, etc.) must be checked against the live docs at https://code.claude.com/docs/en/<page> before being written into a reference file — this skill's entire value is gotcha density, and a wrong gotcha is worse than no gotcha. Require a verbatim sentence from the page and treat "no such sentence exists" as an answer that settles it: a summarizing fetch will invent a version number rather than report an absence. The snapshot in .tmp/docs_claude/ is a research artifact, not a source to verify against — checking claims against it is how six wrong gotchas shipped.
• A reference states the client behavior a model cannot observe at runtime, as the failure shape that changes a placement decision, and not the client-owned value behind it; a number the project itself chooses keeps its reason. SKILL.md's authoring philosophy carries the rule.