Bootstrap a new repository with AGENTS.md, docs structure, and AI collaboration rules from day one.
Consolidate scattered AI prompts, PR comments, and team agreements into versioned governance files.
Add suggested git hooks and CI checks to an existing project as merge proposals rather than overwrites.
Keep AI collaboration rules portable across Claude Code, Codex, and other Agent Skills clients.
Source language is Chinese with an English mirror; install into a target project rather than globally and remove after the files are written.
When you work with an AI coding assistant on a project, the rules for how the assistant should behave often live in scattered places: chat history, one-off prompts, pull request comments, old documents, and verbal team agreements. The kinoward/dayu-harness-skill project aims to turn those one-off prompts into permanent files in your repository, so the rules keep working after any single chat ends. The Chinese name on the README translates roughly to Great Yu Governance Vault, a reference to a figure in Chinese myth who diverted floodwaters rather than blocking them. The recommended way to install it is the Vercel skills CLI, with the command npx skills add kinoward/dayu-harness-skill run inside the project you want to set up. Flags let you target a specific client such as Claude Code, which installs the files under .claude/skills/, or Codex, which installs them under .agents/skills/. After installation you open your AI client and type /dayu-harness, then follow prompts to initialise a new project, merge with existing rules, run a completeness check, or maintain the governance files. The skill itself can be removed afterwards without losing the files it wrote into the project. The files it writes are the long-term product. An AGENTS.md at the repository root acts as the index, with a docs/harness/ directory containing collaboration guides and check scripts. There are suggested git hooks, CI workflow files, commit message and pull request guidelines, plus knowledge directories like docs/design-docs/, docs/troubleshooting/, and docs/references/. The README notes that existing hooks, CI, and lint configuration are presented as merge proposals rather than overwritten directly. The source language is Chinese, with an English mirror in templates.en/ that is kept structurally identical to the Chinese templates/ directory. Interactive prompts at install time show both languages, so an English-speaking user can pick options without reading Chinese. The README cites OpenAI's Harness Engineering concept, an article by Martin Fowler on the same topic, and the open AGENTS.md format as design references. The repository is MIT licensed.
Generated 2026-05-22 · Model: sonnet-4-6 · Verify against the repo before relying on details.