越界 / 不理想

无视 CLAUDE.md 里写好的约定

Ignored the conventions in CLAUDE.md

代码能跑，但不遵守项目规则文件里写明的、或周围代码里明摆着的约定；规则要求的步骤（比如跑某些测试）也被当成“没必要”而跳过。

适用于: Claude Code · Cursor · Codex · Copilot agent

症状

新代码的结构、命名或分层和周围代码不一样；团队划定的边界被越过；CLAUDE.md 或 AGENTS.md 里的某条规则显然没被执行。智能体甚至可能一边复述规则一边不照做。

为什么会这样

被问到为什么还用 PR 时，@alexitcdev 说是为了审查智能体写的代码，因为它们通常会“无视我的 CLAUDE.md 约定”，或者觉得某些测试没必要而跳过。@gibbostudio 经常发现 Opus 和 Fable 写的代码不符合团队设定的边界和代码风格。@_maulik13 抓到过 Opus 偏离项目既有模式；被追问后，它发现自己的假设是错的，随后简化了代码。@WinterSoldier92 说多加的 markdown 文件只是让智能体“挑着用、然后无视”的更多指令。冗长的规则文件会和任务本身争夺注意力，除非有检查去强制，否则没人执行。同一串里，@Spanky4thewin 的建议是：规则写短，把必须做的放在最前面，再加一个 Stop 钩子，让跳过变得不可能。

示例

示意。

text

CLAUDE.md:  "Data access goes through src/repo/*. Run `npm run test:integration` for DB changes."
Agent diff: src/routes/orders.ts  -> raw SQL query inline in the route handler
Agent:      "Integration tests aren't necessary for this small change. Done ✅"

复制即用的修复

按你用的智能体挑选文件。规则片段用英文写，放进任何仓库都能用。

规则

Claude Code：加入 CLAUDE.md

CLAUDE.md

## Follow the project's rules files
- Before editing, read CLAUDE.md / AGENTS.md / .cursor/rules and at least one neighbouring file that
  does something similar. Follow their structure, naming and layering.
- In your summary, list which project rules applied to this change and how you followed each one.
- Required commands in the rules files are not optional. If you think one is unnecessary, ask;
  don't skip it.
- If following a rule is impossible, say so explicitly instead of quietly deviating.

规则

Codex、Copilot 编码智能体等：加入 AGENTS.md

AGENTS.md

## Follow the project's rules files
- Before editing, read CLAUDE.md / AGENTS.md / .cursor/rules and at least one neighbouring file that
  does something similar. Follow their structure, naming and layering.
- In your summary, list which project rules applied to this change and how you followed each one.
- Required commands in the rules files are not optional. If you think one is unnecessary, ask;
  don't skip it.
- If following a rule is impossible, say so explicitly instead of quietly deviating.

Cursor 规则

Cursor：项目规则

.cursor/rules/ignored-project-conventions.mdc

---
description: "AgentFails: Ignored the conventions in CLAUDE.md"
alwaysApply: true
---

## Follow the project's rules files
- Before editing, read CLAUDE.md / AGENTS.md / .cursor/rules and at least one neighbouring file that
  does something similar. Follow their structure, naming and layering.
- In your summary, list which project rules applied to this change and how you followed each one.
- Required commands in the rules files are not optional. If you think one is unnecessary, ask;
  don't skip it.
- If following a rule is impossible, say so explicitly instead of quietly deviating.

钩子

Claude Code Stop 钩子：测试不过就不让智能体收工

Stop 钩子以退出码 2 结束时，Claude 会继续工作并看到 stderr。检查 stop_hook_active 可避免无限循环。把 `npm test` 换成你的测试命令。

.claude/settings.json

{
  "hooks": {
    "Stop": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "python3 -c \"import json,sys,subprocess; d=json.load(sys.stdin); sys.exit(0) if d.get('stop_hook_active') else None; r=subprocess.run('npm test --silent', shell=True, capture_output=True, text=True); sys.stderr.write('Tests are failing. Fix them (without weakening them) before stopping:\\n'+r.stdout[-3000:]+r.stderr[-2000:]) if r.returncode else None; sys.exit(2 if r.returncode else 0)\""
          }
        ]
      }
    ]
  }
}

技能

团队共享的 /verify 技能（Claude Code）

把它提交进仓库，让整个团队（以及每个智能体会话）跑同一套检查。把命令换成你项目里真实的命令。

.claude/skills/verify/SKILL.md

---
name: verify
description: Verify the current change before calling it done. Use before every commit, PR, or "done" message.
---

# Verify

Run every step and paste the real output. Never summarize a step you did not run.

1. **Scope**: run `git diff --stat` and `git status --short`. For every changed file, say which part
   of the request needs it. Any file you cannot justify gets reverted, or listed under
   "Unrequested changes" for the user to decide.
2. **Build and types**: `npm run build` (or the project's build command).
3. **Tests**: `npm test`. Report pass/fail counts. If a test was edited, skipped or deleted,
   say so explicitly with the reason.
4. **Behavior**: exercise the changed behavior once for real (a curl, a CLI run, a screenshot)
   and show what happened.
5. **Report** in this format:
   - Requested: <one line>
   - Changed files: <list, each with its reason>
   - Checks run: <command -> result>
   - Not done / unsure: <list, or "none">

来源与致谢

在 GitHub 上改进这条遇到了、但修复没用？告诉我们