Agent Workflow Kit

Evaluation-first workflow rules for bringing AI coding agents into real software projects.

English Guide · 中文文档 · 中文首页

---

What This Is

Agent Workflow Kit is a tool-neutral guide and agent workflow package for deciding whether a repository needs AI-assisted development rules, and how heavy those rules should be. It is not tied to Codex: any LLM, coding agent, or agent harness can use the guide, templates, and rules.

It helps maintainers answer five practical questions:

• Does this project need an AI workflow at all?
• Should the workflow be minimal, standard, or full?
• When should a spec layer be used?
• What should AI agents verify before claiming work is done?
• Which external actions require explicit maintainer approval?

The central rule is simple:

Evaluate the project first. Then choose the smallest workflow that reduces real risk.

At A Glance

flowchart LR
  Inspect["Inspect repo"]
  Score["Score risk<br/>0-16"]
  Level["Choose level<br/>0 / 1 / 2 / 3"]
  Rules["Copy rules<br/>AGENTS.md"]
  Verify["Verify on<br/>one small change"]

  Inspect --> Score --> Level --> Rules --> Verify

Loading

Why Use It

Problem	What This Project Provides
AI agents jump into code too quickly	A repo inspection and risk scoring flow before changes
Requirements vanish into chat history	Guidance for adding a spec or change-memory layer only when needed
Work is claimed done without evidence	Copy-ready verification and completion rules for AGENTS.md
Tooling stacks overlap and conflict	Conflict rules for specs, plans, reviews, hooks, and external actions
Teams want consistent agent behavior	Agent-facing Skill packages and reusable templates for repeatable evaluations

How To Use The Kit

Path	Best When
Read the English guide or Chinese guide	You want to evaluate a project manually
Copy examples/AGENTS.level-2.md	You want a practical standard workflow starter
Use the skills/agent-workflow-kit package	You want an agent to inspect a repo and produce a workflow decision
Adapt the Markdown rules for Claude, Cursor, Codex, Gemini, or another harness	You want consistent behavior across tools

Quick Start

1. Open the English guide or Chinese guide.
2. Score your project with the 0-16 risk table.
3. Choose the matching workflow level.
4. Copy the smallest useful AGENTS.md template block.
5. Fill in real install, test, lint, build, and browser/E2E commands.
6. Test the workflow on one small change before making it team policy.

Workflow Levels

Level	Use When	Typical Rules
Level 0	AI does not edit code, or the project is disposable	No formal AI workflow
Level 1	Small maintained projects	Basic agent rules and verification commands
Level 2	Most AI-assisted software projects	Basic rules, spec guidance, agent discipline, external-action safety
Level 3	Production, security-sensitive, complex UI, or multi-agent teams	Level 2 plus review gates and stronger verification

Documentation

Document	Purpose
English Guide	Full public guide for maintainers
Chinese Guide	Full Chinese public guide
AGENTS Templates	Copy-ready English AGENTS.md blocks
中文 AGENTS 模板	Copy-ready Chinese Skill reference with English agent blocks
Engineering References	Optional engineering guideline catalog
中文工程规约参考	Chinese reference catalog
Contributing	Contribution and synchronization guide
Changelog	Project change history
Level 2 AGENTS Example	Copy-ready standard workflow starter

Agent Skills And Templates

The skills/ directory contains agent-facing versions of the guide. They are packaged as Codex-compatible Skills, but the instructions are plain Markdown and YAML, so other LLMs, coding agents, and agent harnesses can read and adapt them as project workflow rules.

These packages are optional. If you only want the human guide or AGENTS.md templates, use the documents above.

Skill	Use It For
skills/agent-workflow-kit	English agent Skill package for evaluating a repository, scoring risk, recommending workflow level, and preparing project rules after approval
skills/agent-workflow-kit-zh-cn	Chinese agent Skill package with the same workflow and Chinese-facing instructions

Each skill includes:

• SKILL.md: the agent workflow and safety rules
• agents/openai.yaml: display metadata and default prompt
• references/agents-templates*.md: reusable AGENTS.md template blocks
• references/engineering-references*.md: optional engineering reference catalog

Example prompts:

Use $agent-workflow-kit to evaluate this repository and recommend the right AI workflow level.

使用 $agent-workflow-kit-zh-cn 评估这个仓库，并推荐合适的 AI 工作流级别。

Repository Layout

README.md                         # Project homepage
README.zh-CN.md                   # Chinese project homepage
AGENTS.md                         # AI maintenance rules for this repo
CONTRIBUTING.md                   # Contribution guide
CHANGELOG.md                      # Project change history
.gitattributes                    # GitHub Linguist metadata
docs/guide.md                     # Full English guide
docs/guide.zh-CN.md               # Full Chinese guide
examples/AGENTS.level-2.md        # Copy-ready standard workflow starter
skills/
  agent-workflow-kit/        # English agent Skill package
  agent-workflow-kit-zh-cn/  # Chinese agent Skill package
scripts/check-docs.rb             # Markdown, YAML, skill, and link checks
.github/workflows/docs-check.yml  # GitHub Actions validation

Validation

Run the local checks before publishing documentation changes:

ruby scripts/check-docs.rb

The check validates Markdown fences, YAML syntax, skill structure, and public links.

License

• Guide prose and Skill instructions: CC BY 4.0
• Reusable AGENTS.md templates and example code blocks: CC0 1.0
• Third-party tools and references keep their own licenses.