AGENT SKILL

harness-builder

Analyze a codebase and recommend an agent harness (agents, skills, hooks).

License 许可
Apache-2.0
最近更新
2026-11-13
Stars
★ 0
Forks
1

01 前言 Front Matter

name
harness-builder
description
Analyze a codebase and recommend an agent harness (agents, skills, hooks).
allowed-tools
Read, Grep, Glob, Write, Edit, Bash(git log*), Bash(git status*), Bash(git add*), Bash(ls*)

02 资源文件 2

03 SKILL.md 正文

Harness Builder

Get a harness recommendation tailored to this codebase (agents, skills, hooks, and a CLAUDE.md skeleton) instead of starting from scratch. The harness-builder agent reads the stack, workflows, and domain; you review, trim, and then generate the files. Use it early in a project's setup. Its recommendations are starting points that you are expected to customize.

It needs a git repository (the agent reads history) and a verification command such as a test suite, linter, or build. The agent sorts what it recommends into facts, project-local content, and the generic set of harness-kit skills and agents. This skill writes only the project-local Claude Code files. /harness bootstrap (or $harness bootstrap on Codex) runs the same agent as its first step, then writes the facts and vendors the generic set for both hosts; use it for a full setup, and this skill when you want the recommendation and a Claude-side starting point.

What the agent tends to recommend, and how to read it, is in recommendations.md.

1. Run the agent

Dispatch the harness-builder agent on the current codebase. Pass it the absolute paths of the harness skill's catalogs so it recommends from them rather than from memory: external-plugins.md, architecture-patterns.md, hooks-reference.md, and release-please-reference.md. The harness skill sits beside this one, in <skill-dir>/../harness/, in the plugin and in a vendored snapshot alike. Also pass the generic names it can choose from: the skill folders beside this one and the plugin's agents.

Done when the agent has returned its report and the report does not say the catalogs were missing; if it does, dispatch it again with the paths. When the report misreads the stack, use the questions under "Stack detection" in recommendations.md and fold the answers in.

2. Review the recommendations

The report covers:

  • Stack: language, frameworks, and the verification command.
  • Facts: the verification command as [facts] verification_command, then the prose for AGENTS.md's ## Project facts section.
  • Agents: the analytical work to hand off, split into generic (code-reviewer, verifier) and project-local (test-writer, domain-advisor). Judge roles and model tiers against "Agents" in recommendations.md.
  • Skills: the workflows to provide, split into the generic set (always /commit, plus GitHub and frontend tools when detected) and project-local skills. "Skills" in recommendations.md covers when a need is a fact and when it is a local skill.
  • Hooks: formatters, validators, and guidance for auto-fixing. "Hooks" in recommendations.md has a working example; the reasoning behind the stages is in <skill-dir>/../harness/hooks-patterns.md.
  • Domain knowledge: entity types, ownership, the auth and session system, and the side effects of mutating core entities.

Prefer the generic skills (/commit, /review-pr) the agent points to, with the project's facts stated for them, over generating new ones. Done when every recommendation is either accepted or explained away to the user.

3. Customize

Remove the agents and skills the project doesn't need, add project-specific agents or constraints, and put the project's actual domain rules into CLAUDE.md in place of boilerplate. Done when the user approves the trimmed list.

4. Generate the files

After approval, generate the project-local content only:

  • .claude/agents/*.md, one per project-local agent, with its tool permissions and domain context.
  • .claude/skills/*/SKILL.md, one per project-local skill, with allowed-tools and a description that says what it does and when.
  • CLAUDE.md, the harness documentation, with the facts.
  • .gitignore updates.

The generic set is never copied by hand; /harness bootstrap vendors it.

Then run the project's verification command to confirm the setup works. On a Nix flake, git add the new files first, because flakes only see tracked files. Done when every approved file exists and the verification command passes.

Related

  • /harness: audits harness quality, and bootstraps both hosts.
  • /skill-writer: well-structured skills.

04 同仓库更多 Skills

dougborg
dougborg/harness-kit

teach

Teach yourself a topic over several sessions, with lessons, references, and a record of what you've learned.

教师与教育
★ 0
dougborg
dougborg/harness-kit

svg-logo-designer

Generate SVG logos with multiple concepts, layouts, and color variations.

图像与多媒体
★ 0
dougborg
dougborg/harness-kit

documentation-writer

Writes and revises human-facing documentation — READMEs, contributor and setup guides, reference pages, architecture notes, CLAUDE.md, and changelog entries. Covers picking the right document type for the question being asked, structuring for skimming, writing examples that actually run, and keeping docs from going stale. Use when creating or rewriting docs, when a README has drifted from the code, or when deciding where a piece of information belongs. For SKILL.md and agent files, use /skill-writer instead.

内容创作者写作与文案
★ 0
dougborg
dougborg/harness-kit

adhd-mode

Shapes every response for a reader with ADHD: the next action first, numbered steps, progress restated each turn, tangents held back, concrete time estimates, wins made visible. Stays on for the rest of the session until the reader says "stop adhd mode". On Claude Code, the harness-kit:adhd output style does the same without invoking this skill.

★ 0
dougborg
dougborg/harness-kit

wait-what

That last message didn't land: re-explain it with the missing context.

★ 0
dougborg
dougborg/harness-kit

to-questionnaire

Turn a decision you can't make alone into a questionnaire for the one person who can answer it, filled in async or together in a meeting.

★ 0