Skip to content

Loop Engineering 专题提示词

来源：loops!；抓取日期：2026-06-29。loops! 将一次性提示词整理为“触发器 → 执行步骤 → 反馈检查 → 退出条件”的闭环工作流，让 Coding Agent 能自我节奏化地反复执行，直到达到可验证的完成状态。

如何使用这些 Loop

  1. 选择与你当前任务最接近的 Loop。
  2. 复制对应的 Kickoff Prompt 到 Cursor、Claude Code、Codex、Gemini CLI 或其他 Coding Agent。
  3. 按照提示词中的检查命令在每轮迭代后验证结果。
  4. 只有当退出条件满足时才停止；如果多轮后仍无法通过，应报告阻塞点，而不是修改检查命令或绕过测试。

Loop 速览

Loop 场景 触发方式 检查命令 退出条件
Ship PR Until Green 开分支、提交 PR、修复 CI 直到可合并 手动 gh pr checks PR 检查全部成功
CI Failure Watcher 定时查看当前分支 CI，失败就定位并修复 定时 gh run list --branch $(git branch --show-current) --limit 1 最新 CI run 成功
Independent Verifier Pass 独立验证构建、Lint、测试，不依赖实现者自述 手动 npm run build && npm run lint && npm test 所有验证命令返回 0
Post-Edit Test Guard 每批文件编辑后运行相关测试，提前发现回归 事件 npm test -- --findRelatedTests <edited files> 相关测试通过
Pre-Commit Guard 提交前运行测试，阻止红灯提交 事件 npm test 提交前测试通过
Deploy Verification Loop 部署后定时检查健康检查与冒烟端点 定时 curl -fsS <your-health-url> 所有端点成功
Guardrails Learning Loop 同类失败重复出现时，把经验写入 guardrails 手动 npm test && npm run lint 测试和 Lint 通过，且不重复旧失败模式
Flaky Test Triage 反复运行失败测试，区分 flaky 与真实回归 手动 npm test -- --testPathPattern=<failing-suite> 失败均已分类，真实回归已修复或延期
Docs Sync After Edits 代码变更后同步 README、API 文档和注释 手动 git diff main...HEAD --name-only 受影响文档已更新并验证
npm Audit Fix Loop 逐个修复高危或严重 npm audit 漏洞 手动 npm audit --audit-level=high && npm test 无 high/critical 漏洞
A11y Audit Until Clean 对变更 UI 运行无障碍审计，修复严重问题 手动 npm run test:a11y 无障碍审计退出码为 0

通用防作弊规则

这些 Loop 的关键价值在于验证驱动，因此 Agent 必须遵守以下规则：

  • 不要修改检查命令或退出条件来制造成功。
  • 不要跳过、禁用或绕过检查。
  • 如果多轮迭代后仍然卡住，停止并报告阻塞点。
  • 涉及测试的 Loop 中，不要削弱、删除或跳过测试来让套件变绿。
  • 不要把真实断言替换成永远通过的空断言。
  • 优先修复生产代码，而不是为了变绿而随意改测试。

Ship PR Until Green

适用场景： 实现一项变更，创建或更新 PR，然后持续修复 CI，直到 PR 处于可合并状态。

工作步骤：

  1. 实现变更并在本地运行测试。
  2. 使用清晰的提交信息提交并推送分支。
  3. 创建或更新 PR，检查 PR 状态。
  4. 如果 CI 失败，读取日志，修复根因，本地验证后再推送。

Kickoff Prompt：

Start the "Ship PR Until Green" loop.

Goal: PR is open with all CI checks passing
Max iterations: 10
Between iterations run: gh pr checks
Exit when: all PR checks are success

Step 1: Implement the change, test locally, push, open PR, and fix CI until green.

Self-pace this loop. After each iteration, run the check command, read the output, and only continue if the exit condition is not met. Stop when the exit condition passes or max iterations is reached. Give a short status update each pass.

CI Failure Watcher

适用场景： 当前分支已有 CI 运行，需要定时轮询，发现红灯后自动分析日志并修复。

工作步骤：

  1. 使用 GitHub CLI 检查当前分支最新 CI run。
  2. 如果 CI 失败，阅读日志并定位根因。
  3. 本地复现失败的 CI 步骤，应用最小修复。
  4. 必要时推送修复并继续轮询，直到最新 CI 成功。

Kickoff Prompt：

Start the "CI Failure Watcher" loop.

Goal: latest CI run on this branch is green.
Max iterations: 12.
Between iterations run: gh run list --branch $(git branch --show-current) --limit 1
Exit when: latest run conclusion is success.

Step 1: Check CI status. If failed, read logs, fix root cause, verify locally, and push if needed.

Independent Verifier Pass

适用场景： 实现者声称完成后，让一个独立验证轮次只相信命令输出，不相信口头结论。

工作步骤：

  1. 作为独立验证者运行 build、lint 和测试。
  2. 汇总所有失败检查，包含文件路径与错误摘录。
  3. 如果可以，修复失败；否则把简洁的失败报告交回实现者。

Kickoff Prompt：

Start the "Independent Verifier Pass" loop.

Goal: build, lint, and tests pass under independent verification
Max iterations: 8
Between iterations run: npm run build && npm run lint && npm test
Exit when: all verifier commands exit 0

Step 1: Run build, lint, and tests as a verifier. Trust only command output, not prior claims.

Self-pace this loop. After each iteration, run the check command, read the output, and only continue if the exit condition is not met. Stop when the exit condition passes or max iterations is reached. Give a short status update each pass.

Post-Edit Test Guard

适用场景： 每次完成一批文件编辑后，先运行与改动相关的最小测试集合，防止回归继续扩大。

工作步骤：

  1. 识别上一轮改动过的源文件。
  2. 运行与这些文件相关的测试子集。
  3. 如果测试失败，先修复失败，再继续做更多改动。

Kickoff Prompt：

Install and run the "Post-Edit Test Guard" loop.

Goal: after each batch of file edits, related tests must pass before continuing.
Between iterations run: npm test -- --findRelatedTests <edited files>
Exit when: related tests exit 0.

Step 1: After edits, run related tests. If they fail, fix before making more changes.

Pre-Commit Guard

适用场景： 在执行 git commit 前强制运行测试，避免把红灯状态提交进历史。

工作步骤：

  1. 检测提交意图。
  2. 提交前运行完整测试套件。
  3. 测试失败则阻止提交并先修复；测试通过才继续提交。

Kickoff Prompt：

Install and run the "Pre-Commit Guard" loop.

Goal: block git commits when tests are failing.
Between iterations run: npm test
Exit when: tests exit 0 before each commit.

Step 1: Before any git commit, run tests. Fix failures before committing.

Deploy Verification Loop

适用场景： 部署后持续访问健康检查和冒烟测试端点，直到所有端点返回健康结果。

工作步骤：

  1. 请求配置好的健康检查或冒烟测试 URL，记录状态码和响应体。
  2. 如果任一端点失败，检查部署日志、环境变量和数据库迁移等。
  3. 应用最小修复或做出回滚/升级判断，然后重新验证。

Kickoff Prompt：

/loop 15m Start the "Deploy Verification Loop".
Goal: all post-deploy health and smoke endpoints return success.
Max iterations: 8.
Between iterations run: curl -fsS <your-health-url>
Exit when: every configured endpoint succeeds.
Step 1: Hit health/smoke URLs. If any fail, inspect deploy logs and fix or escalate.

Guardrails Learning Loop

适用场景： 当测试或 Lint 以相同方式失败两次时，把可复用的“护栏标记”写入 .ralph/guardrails.md，让下一轮迭代避免重复犯错。

工作步骤：

  1. 读取 .ralph/guardrails.md 中已有的护栏。
  2. 运行测试和 Lint，观察失败模式是否重复。
  3. 如果同类失败第二次出现，先追加一条 guardrail sign，再修复问题。
  4. 每轮修复后重新运行检查，直到所有检查通过。

Kickoff Prompt：

Start the "Guardrails Learning Loop" loop.

Goal: tests and lint pass without repeating prior failure patterns
Max iterations: 12
Between iterations run: npm test && npm run lint
Exit when: all checks pass

Step 1: Read .ralph/guardrails.md, run checks, and if a failure repeats, add a sign before fixing.

Self-pace this loop. After each iteration, run the check command, read the output, and only continue if the exit condition is not met. Stop when the exit condition passes or max iterations is reached. Give a short status update each pass.

Guardrails (do not skip):
- Do not modify the check command or exit criteria to force success
- Do not skip, disable, or bypass checks to pass the exit condition
- If stuck after several iterations, stop and report blockers instead of gaming metrics

Flaky Test Triage

适用场景： 某个测试套件失败但不确定是否为 flaky，需要反复运行并分类，避免把不稳定测试与真实回归混在一起处理。

工作步骤：

  1. 多次运行失败的测试套件。
  2. 将每个失败分类为 flaky 或真实回归。
  3. 只修复已确认的真实回归。
  4. 记录 flaky 行为，必要时明确延期处理。

Kickoff Prompt：

Start the "Flaky Test Triage" loop.

Goal: classify failing tests as flaky vs real and fix only real regressions
Max iterations: 5
Between iterations run: npm test -- --testPathPattern=<failing-suite>
Exit when: every failure is classified and real regressions are fixed or explicitly deferred

Step 1: Run the failing suite multiple times. Classify each failure, fix real ones, and document flaky behavior.

Self-pace this loop. After each iteration, run the check command, read the output, and only continue if the exit condition is not met. Stop when the exit condition passes or max iterations is reached. Give a short status update each pass.

Guardrails (do not skip):
- Do not modify the check command or exit criteria to force success
- Do not skip, disable, or bypass checks to pass the exit condition
- If stuck after several iterations, stop and report blockers instead of gaming metrics
- Do not weaken, delete, or skip tests to make the suite pass
- Do not replace real assertions with trivial always-pass tests
- Prefer fixing production code over patching tests to go green

Docs Sync After Edits

适用场景： 代码修改后需要同步 README、API reference、示例和行内注释，防止文档与实现脱节。

工作步骤：

  1. 查看当前分支相对 main 的改动文件。
  2. 找出受影响的 README、API 文档、示例和行内注释。
  3. 更新过期内容，并核对描述与当前代码一致。
  4. 重新检查 diff，确认相关文档已覆盖。

Kickoff Prompt：

Start the "Docs Sync After Edits" loop.

Goal: documentation matches the current code changes
Max iterations: 3
Between iterations run: git diff main...HEAD --name-only
Exit when: all affected docs are updated and verified

Step 1: Review the diff, find stale docs, update them, and verify accuracy.

Self-pace this loop. After each iteration, run the check command, read the output, and only continue if the exit condition is not met. Stop when the exit condition passes or max iterations is reached. Give a short status update each pass.

Guardrails (do not skip):
- Do not modify the check command or exit criteria to force success
- Do not skip, disable, or bypass checks to pass the exit condition
- If stuck after several iterations, stop and report blockers instead of gaming metrics

npm Audit Fix Loop

适用场景： npm audit 发现 high 或 critical 漏洞，需要逐个选择安全修复并验证测试，而不是盲目执行破坏性 npm audit fix --force。

工作步骤：

  1. 选择一个 high/critical advisory。
  2. 应用最安全、影响最小的修复方案。
  3. 运行 audit 和测试验证。
  4. 通过后再处理下一个 advisory。

Kickoff Prompt：

Start the "npm Audit Fix Loop" loop.

Goal: no high or critical npm audit vulnerabilities
Max iterations: 10
Between iterations run: npm audit --audit-level=high && npm test
Exit when: npm audit reports no high/critical issues

Step 1: Pick one high/critical advisory, apply the safest fix, run tests, and repeat.

Self-pace this loop. After each iteration, run the check command, read the output, and only continue if the exit condition is not met. Stop when the exit condition passes or max iterations is reached. Give a short status update each pass.

Guardrails (do not skip):
- Do not modify the check command or exit criteria to force success
- Do not skip, disable, or bypass checks to pass the exit condition
- If stuck after several iterations, stop and report blockers instead of gaming metrics

A11y Audit Until Clean

适用场景： UI 路由有变更后，对变更页面运行自动化无障碍审计，优先修复键盘导航和屏幕阅读器相关的严重问题。

工作步骤：

  1. 对变更路由运行无障碍审计。
  2. 逐项修复审计发现的 violation。
  3. 优先处理键盘可达性、语义结构和屏幕阅读器体验。
  4. 重复审计，直到命令退出码为 0。

Kickoff Prompt：

Start the "A11y Audit Until Clean" loop.

Goal: zero serious accessibility violations on changed UI
Max iterations: 8
Between iterations run: npm run test:a11y
Exit when: a11y audit exits 0

Step 1: Run the a11y audit on changed routes. Fix each violation, prioritizing keyboard and screen reader issues.

Self-pace this loop. After each iteration, run the check command, read the output, and only continue if the exit condition is not met. Stop when the exit condition passes or max iterations is reached. Give a short status update each pass.

Guardrails (do not skip):
- Do not modify the check command or exit criteria to force success
- Do not skip, disable, or bypass checks to pass the exit condition
- If stuck after several iterations, stop and report blockers instead of gaming metrics

设计自己的 Loop Prompt 模板

当你把普通提示词改写为 Loop Engineering 提示词时，可以使用下面的结构：

Start the "<Loop Name>" loop.

Goal: <用可验证语言描述目标>
Max iterations: <最大迭代次数>
Between iterations run: <每轮必须运行的检查命令>
Exit when: <明确、客观、可观察的退出条件>

Step 1: <第一步行动>
Step 2: <第二步行动>
Step 3: <验证、修复或升级策略>

Self-pace this loop. After each iteration, run the check command, read the output, and only continue if the exit condition is not met. Stop when the exit condition passes or max iterations is reached. Give a short status update each pass.