Token导航 LogoToken导航

Agent Skill · 研究检索

review-driven-development(审查驱动开发)

review-driven-development 用于查找、检索和筛选相关信息,适合在 Codex、Claude、Cursor、Gemini CLI 中需要根据关键词、任务场景或来源线索快速定位候选结果时使用。可结合来源仓库、安装命令和原始 README 继续核验具体用法。安装前建议确认权限范围、维护状态,以及是否会触发联网、命令执行或文件读写。

review-driven-development 是研究检索类 Agent Skill,可用宿主为 Codex、Claude、Cursor、Gemini CLI,通过GitHub 仓库获取,权限为需要联网;许可证信息还不明确,第三方扫描有提醒项,按安装量在研究检索类的 73,675 个 Skill 里排第 67,168。

总安装 250周安装 10GitHub Stars 28下载 81核验 2026-05-01

安装说明

本站只整理中文说明,不托管安装包,也不代用户安装。
GitHub
许可证
unknown

安装方式

通过对话安装

复制提示词发给支持本地命令或 Skills 的 AI 助手,先确认命令和权限,再让它执行。

请帮我安装这个 Agent Skill:review-driven-development(审查驱动开发)
来源仓库:https://github.com/yvictor/skills
仓库路径:skills/review-driven-development
安装命令:
npx skills add https://github.com/yvictor/skills --skill review-driven-development
安装前请先检查当前环境是否支持对应 CLI,并向我确认将要执行的命令、安装目录、联网范围和文件读写权限;确认后再执行。

命令行安装

复制命令到本机终端执行。该命令会通过 npx skills 从第三方来源获取 Skill;本站只展示命令,不托管安装包,也不自动执行。

skills.shnpx skills
npx skills add https://github.com/yvictor/skills --skill review-driven-development

这个 Skill 是做什么的

研究检索类 Skill 主要解决「找资料、核对事实、把来源整理清楚」的问题,通常会让助手先检索再总结。

review-driven-development 用于查找、检索和筛选相关信息。适合在 Codex、Claude、Cursor、Gemini CLI 中根据关键词或任务场景快速定位候选结果。通过 npx skills add 命令从指定 GitHub 仓库安装使用。

安装前需确认权限范围、维护状态,注意是否触发联网或文件操作。建议结合原始 README 核验具体用法和功能边界。

来源仓库内路径:skills/review-driven-development

上手步骤

4 步
  1. 1

    确认自己的宿主工具

    公开资料标注支持 Codex、Claude、Cursor、Gemini CLI;先确认自己用的是其中之一,再按对应工具的 Skills 目录放置。

  2. 2

    拿到安装命令

    可以直接复制的命令是 npx skills add https://github.com/yvictor/skills --skill review-driven-developme…;命令里的占位符换成自己的路径或密钥再执行。

  3. 3

    安装后重启宿主

    命令执行完要完全退出宿主工具再打开,让新装的 Skill 被重新扫描到,否则列表里可能看不到。

  4. 4

    按最小权限试用

    它需要需要联网,先在测试目录或测试账号里跑,确认没问题再放到正式项目。

常见问题

6 条

review-driven-development 怎么安装?

在终端执行 npx skills add https://github.com/yvictor/skills --skill review-driven-development。装完完全退出宿主工具再打开,让它重新扫描本地 Skill。

review-driven-development 支持哪些 AI 工具?

公开资料标注支持 Codex、Claude、Cursor、Gemini CLI。其他工具只要能读取本地 Skill 目录,一般也能手动放进去,但要自己确认触发方式是否一致。

review-driven-development 需要什么权限?

权限级别是需要联网。该 Skill 可能需要联网访问来源站点、仓库或外部 API;具体网络访问范围需要结合源码和 README 复核。

review-driven-development 是免费的吗?

公开资料没有标注明确许可证,收费和商用限制都无法确认。正式使用前请到来源仓库核对授权说明。

review-driven-development 用的人多吗?

公开数据是累计安装 250 次,近一周安装 10 次,下载 81 次,GitHub 28 星。这些数字来自第三方目录统计,只能当作热度参考,不代表质量或安全性。

review-driven-development 安全吗?

第三方扫描有提醒项,说明存在需要人工确认的行为。安装前请打开源码看清它读了什么、联网到哪里。

值得选的地方和要先想清楚的地方

值得选的地方

  • 同时适配 4 种工具,换客户端不用重新找替代品。

先想清楚的地方

  • 许可证没有明确标注,能不能商用、能不能再分发都不确定。
  • 第三方扫描存在提醒项或未通过项,需要人工逐条确认。
  • 第三方目录整理,版本更新和行为细节以来源仓库为准。

适不适合你

更适合

  • 已经在用 Codex,想直接扩展它能力的人。
  • 需要处理研究检索类任务、又不想每次都手写提示词的人。

不太适合

  • 需要明确授权才能用到公司项目里的场景。
  • 需要厂商级技术支持和 SLA 的生产系统。

研究检索类 Skill 对比

6 个同类
Skill总安装GitHub 星权限许可证扫描
find-skills

寻找技能 · Codex、Claude

31,199,81316,486操作浏览器MIT有提醒
self-improving-agent

自我改进 Agent · OpenClaw

16,464,0023,396接触敏感数据MIT-0通过
soultrace

灵魂追踪 · Codex、Claude

7,588,94412需要联网未明确有提醒
self-improving

自我完善 · OpenClaw

6,829,2841,046需要联网MIT-0通过
microsoft-foundry

微软代工厂 · Codex、Claude

6,614,144764调用外部服务未明确有提醒
review-driven-development

审查驱动开发 · 本页

25028需要联网未明确有提醒

按安装量取同分类前几名;数字来自第三方目录统计,只作热度参考,不代表质量结论。

SKILL.md 原文

Review-Driven Development

A complete issue-to-merge workflow where the iterative review cycle is the core driver. The PR/MR doesn't just get created — it gets reviewed N rounds with fixes until the verdict is APPROVE.

Announce at start: "Using review-driven-development to implement issue #NN through the full review cycle."

Overview

Issue → Research → Plan → Worktree → Implement → Test → PR/MR → Review ⟲ Fix → APPROVE → Merge
                                                                 ↑_____________↩ (repeat until clean)

The review cycle (Phase 5) is what separates this from a simple "code and push" workflow. Every PR/MR goes through at least one structured review round, with findings posted directly to the PR/MR. Issues get fixed, re-reviewed, and the cycle repeats until the verdict is APPROVE.


Phase 1: Issue Research

When the user provides an issue number:

  1. Fetch issue details — auto-detect platform:

- GitHub: gh issue view <NN> --json title,body,labels,assignees - GitLab: glab issue view <NN> - If neither works, ask the user for the issue URL

  1. Research the codebase — launch Explore agents to understand:

- What code is involved - What patterns exist in the codebase - What reference repos or docs are relevant (check CLAUDE.md, memory)

  1. Enter plan mode to present the implementation plan

Phase 2: Plan + Worktree

The plan MUST include:

  • Worktree setup: directory path, branch name (read CLAUDE.md/memory for naming conventions)
  • Files to modify: list the key files with what changes each needs
  • Existing code to reuse: functions, patterns, utilities found during research
  • Test strategy: what commands to run (read from CLAUDE.md)
  • Verification: how to confirm the changes work
  • Workflow phases: structure the plan around this skill's 6 phases (Issue Research → Plan + Worktree → Implementation → Commit + PR/MR → Review Cycle → Merge + Cleanup), so the plan serves as both an implementation guide and a workflow tracker

After user approves the plan:

  1. Create worktree + branch
  2. Rebase onto latest default branch (origin/master or origin/main)
  3. Begin implementation

Phase 3: Implementation

Execute the plan following project-specific patterns from CLAUDE.md and memory.

  • Follow existing code patterns — read similar files first
  • Run project-specific build/test commands (from CLAUDE.md)
  • Run real examples or integration tests if applicable
  • This phase composes with other skills:

- If TDD skill is active → write tests first, then implementation - If systematic-debugging skill is active → use it when hitting issues - If brainstorming skill is active → use it for design decisions

Phase 4: Commit + PR/MR

  1. Rebase onto latest default branch (git fetch origin && git rebase origin/master or origin/main). If conflicts arise, enter plan mode to analyze the conflicts and plan the resolution strategy before proceeding.
  2. Stage specific files — never git add -A or git add.
  3. Commit with descriptive message
  4. Push to remote with -u flag (use --force-with-lease if rebased)
  5. Create PR/MR:

- GitHub: gh pr create --title "..." --body-file /tmp/pr_body.md - GitLab: glab mr create --title "..." --description "$(cat /tmp/pr_body.md)"

  1. PR/MR body format:
## Summary
- Bullet points describing what changed and why

Closes #NN

## Test plan
- [x] Test results summary
- [x] Verification steps completed

Phase 5: Review Cycle (THE CORE)

This is the heart of review-driven development. The PR/MR enters an iterative review loop.

Step 1: Choose review strategy

Ask the user which review strategy to use (only on first review round):

StrategyReviewersWhen to use
SingleCurrent agent (code-reviewer subagent)Quick iterations, small changes
DoubleAgent + codex exec "..."Important features, want second perspective
TripleAgent + codex exec "..." + claude -p "..."Critical changes, maximum coverage

Step 2: Execute reviews

For each reviewer, provide the full PR/MR diff and ask them to check:

  • Pattern compliance with the codebase (compare against similar existing files)
  • Error handling consistency
  • Test coverage
  • Type safety and correctness
  • Any issues categorized as:

- B (Bug/Important) — must fix before merge - S (Suggestion) — nice to have - N (Note) — informational only

External reviewers:

# Codex review
codex exec "Review PR #NN for <repo>. Run 'git diff master...HEAD' to see the diff. Check: 1) pattern compliance 2) error handling 3) test coverage 4) correctness. Write findings."

# Claude review
claude -p "Review the git diff for PR #NN. $(git diff master...HEAD). Check pattern compliance, error handling, test coverage. Write findings with B/S/N categorization and a verdict."

Step 3: Post combined review

Post the combined review findings to the PR/MR:

# GitHub
gh api repos/{owner}/{repo}/pulls/{NN}/reviews -X POST -f body="$(cat /tmp/review.md)" -f event="COMMENT"

# GitLab
glab mr note {NN} -m "$(cat /tmp/review.md)"

The review comment MUST include:

  1. Pattern compliance checklist — table of checks with PASS/FAIL
  2. Issues — categorized as B (bug), S (suggestion), N (note)
  3. Verdict — APPROVE or REQUEST_CHANGES

Step 4: Fix issues

Ask the user which issue levels to address (B/S/N or any combination). Then:

  1. Enter plan mode — the plan should clearly state:

- Current phase: Phase 5: Review Cycle - Current round: Round N (increment from previous round) - Issues to fix: list each issue the user chose to address, with its level (B/S/N) - Fix strategy: what changes to make for each issue - Re-review plan: after fixes, run the same review strategy again (back to Step 2)

  1. Implement fixes
  2. Re-run tests
  3. Commit + push (new commit, don't amend)
  4. Update PR/MR body to reflect current state (updated test counts, fixed issues, etc.)

Step 5: Re-review

Rebase onto latest default branch if needed (enter plan mode if conflicts arise), then run the same review strategy again on the updated code. Post the new review to the PR/MR.

Step 6: Repeat

Continue the fix → re-review loop until the verdict is APPROVE from all reviewers.

Phase 6: Merge + Cleanup

This phase REQUIRES explicit user confirmation. Do NOT auto-merge. Wait for the user to say "merge", "ok merge", "ship it", etc.

Once confirmed:

  1. Merge:

- GitHub: gh pr merge <NN> --merge (or --squash/--rebase per CLAUDE.md) - GitLab: glab mr merge <NN>

  1. Cleanup:

- Remove worktree: git worktree remove <path> - Delete local branch: git branch -d <branch>

  1. Confirm: show the merge commit on the default branch

Principles

  • CLAUDE.md first — always read project instructions before starting. Build commands, test commands, naming conventions, merge strategy — it's all there.
  • Memory — check auto-memory for user preferences and project context from prior sessions.
  • Platform agnostic — support both GitHub (gh) and GitLab (glab). Auto-detect from git remote.
  • Review is non-negotiable — every PR/MR gets at least one structured review round with findings posted to the PR/MR itself.
  • Composable — this skill handles the lifecycle; other skills handle methodology (TDD, debugging, brainstorming).
  • Only Phase 6 blocks on user — all other phases proceed autonomously. Phase 6 (merge+cleanup) waits for explicit user confirmation.

平台分布

Codex

36.2%
按下载量换算29

Claude

32.25%
按下载量换算26

Cursor

16.91%
按下载量换算14

Gemini CLI

8.4%
按下载量换算7

安全审计

Gen Agent Trust Hub

通过

Socket

可疑

Snyk

可疑

权限和风险

需要联网

该 Skill 可能需要联网访问来源站点、仓库或外部 API;具体网络访问范围需要结合源码和 README 复核。

安装前确认

本站仅展示第三方公开信息,不托管安装包,不提供自动安装或运行环境。安装前应自行审查源码、依赖和命令行为。来源安全扫描存在提醒或未通过项,安装前请自行复核。目前只有一个公开来源,建议同时核对源仓库。

继续浏览同类 Skills