harness-dev

Use when writing code examples inside a rules or docs file that agents will copy verbatim — CLAUDE.md, .claude/rules/*.md, a SKILL.md, a style guide, a contributing doc. The linter never sees these snippets, so a good/bad example can teach a shape the repo's own lint job rejects. Check the enabled linter set before writing the snippet.

Use when writing or fixing a `<router>:help` slash command in a plugin under harness/plugins/ — a command whose only job is to print a static help table/guide. Trigger on "add a help command", "write a help command", "the help command prints junk / shows load summary instead of the table", "fix code:help/work:help".

Use when adding an eval suite, test fixtures, or a grading runner for a skill that lives in a plugin under harness/plugins/. Trigger on "write evals for this skill", "add a rubric", "grade the gate", or when choosing where cases.jsonl and run.sh go.

Use when the user runs "/prompt-dry-run <prompt>" or asks to draft/preview/dry-run how a prompt would be handled — reflect back the agent's reading of the prompt (intent, planned actions, limits) BEFORE acting, so misalignment surfaces first. Trigger on "draft this", "what would you do with", "before you act, tell me your plan", "how do you read this".

Rewrite a skill, an AGENTS.md/CLAUDE.md, or a doc corpus into the agent-writing house style — cut duplication across files, cut dead weight inside each file, then reshape into steps plus disclosed reference. Use when the user asks to prune, dedup, compress, or restyle docs or skills, or says files repeat the same information.

Set up a repository's base documentation and Claude configuration — README.md, CLAUDE.md, a per-repo deploy skill, a per-repo adr skill, and .claude/settings.json. Use for "set up this repo", "add repo settings", "bootstrap README and CLAUDE.md", "add a deploy skill to this repo", "set up ADRs here".

Use when writing or editing an always-loaded rules file — a CLAUDE.md, or any other file the harness injects into every session — or when writing a cross-skill pointer inside one. Owns the two conventions those files carry; general prose shape lives in `text-style`.

Audit why a Claude Code session was slow or expensive, and rank fixes onto existing harness levers. Triggers on "make sessions faster", "why did that session burn so many tokens", "where did the context go", "analyse this transcript", "too many turns per prompt", or any ask to profile a .jsonl under ~/.claude/projects.

Router for authoring a Claude Code skill. Use when the user wants to "build a skill", "write a skill", "create a skill", "author a SKILL.md", or build a specific shape — a "workflow" (step-by-step guide), a "loop" (repeat a flow), an "instruction" (rules / reference), or a "router" (dispatch table over branches). Names each shape and routes to its guide.

Shape agent-facing markdown into the house style — steps plus disclosed reference, a checkable completion criterion per step, leading words instead of restatements, positive targets instead of bans, and a description that fires as a context pointer. Load it when writing or rewriting a SKILL.md, an agent file, a command file, a `references/` doc, or an AGENTS.md/CLAUDE.md, when the `prune-text` skill reaches its reshape phase, and when judging whether a file reads as the shape its reader needs.