Pluto｜AI Agent 实测 · Hands-on

每周分享能跑通的 Claude Code / Codex / Cursor Agent 工作流，附截图实测，不吹不黑 🛠
Hands-on AI coding agent workflows, tested with screenshots.

[Pluto｜AI Agent 实测 · Hands-on @Mars_Peakul]

Packaged the review prompts I keep reusing: full PR, security, perf, tests, API, migrations, a11y, deps. Ready to paste into Cursor, Claude Code, ChatGPT or Copilot Chat.
AI Code Review Prompt Pack, $19 → https://t.co/aal5Dhz7ig

[Pluto｜AI Agent 实测 · Hands-on @Mars_Peakul]

Sneakiest AI PR pattern: CI goes green because a test got loosened.
Always review the test diff on purpose. Ask:
"For each behavior change, which test covers it? Was any assertion removed or relaxed?"

[Pluto｜AI Agent 实测 · Hands-on @Mars_Peakul]

Essays aren't merge gates.
My review prompt ends with an output contract:
• Must fix / Should fix / Nit
• every item cites path:line
• final verdict: APPROVE | REQUEST_CHANGES | NEEDS_HUMAN
No contract, no gate.

[Pluto｜AI Agent 实测 · Hands-on @Mars_Peakul]

Bonus: the map itself tells you which human reviewer to tag.

[Pluto｜AI Agent 实测 · Hands-on @Mars_Peakul]

Stop asking AI to review a whole PR evenly. Ask for a risk map first 🧵

[Pluto｜AI Agent 实测 · Hands-on @Mars_Peakul]

Then deep-dive only the top 2–3. A CSS tweak and a token check don't deserve the same attention.

[Pluto｜AI Agent 实测 · Hands-on @Mars_Peakul]

"Review this PR" is how you get 6 polite nits and a fake LGTM.

The model doesn't know what the change is supposed to do, or what counts as blocking.

Give it both: the intent in one line, plus a severity scale. The review gets sharper immediately.

[Pluto｜AI Agent 实测 · Hands-on @Mars_Peakul]

@ScarletWave7 对，支付那波就是先各自圈地、再被监管推着互联互通。Agent 这边可能会快一点：只要有一家大厂先开放授权接口，用户体验差距一拉开，其他家不跟就会被用户用脚投票。

[Pluto｜AI Agent 实测 · Hands-on @Mars_Peakul]

@shtse8 Same trap bit me. What fixed it: keep AGENTS.md as the single source and shrink CLAUDE.md to one line, @AGENTS.md, so Claude Code imports it instead of skipping it. Then run /memory once to confirm what actually loaded.

[Pluto｜AI Agent 实测 · Hands-on @Mars_Peakul]

@fakeroo8 pre-commit 里直接 fail 这个好。^\.env(\..+)?$ 其实已经能盖住 .env.local.backup，但 env.bak、.env~ 这种编辑器备份会漏，我准备再加一层按内容扫 key= / token= 的兜底。

[Pluto｜AI Agent 实测 · Hands-on @Mars_Peakul]

只能留一个 AI coding 工具，你选哪个？

A. Cursor（IDE 内体验最顺）
B. Claude Code（终端 + 长任务）
C. Codex（OpenAI 生态）
D. 都不选，开源 + 自己的 API

评论区说说理由，我挑最有意思的几个做一期对比教程👇

[Pluto｜AI Agent 实测 · Hands-on @Mars_Peakul]

@ScarletWave7 客气～其实技术上都能做，难的是账号、数据和支付各家不互通，Agent 每跨一个 App 就得重新授权一次。哪天有个统一的授权层，国内这类产品很快就会冒出来。

[Pluto｜AI Agent 实测 · Hands-on @Mars_Peakul]

文字版，可以直接贴进 AGENTS.md / CLAUDE.md：
1. 错误路径：断网、空数据、超时各补一个测试
2. 密钥：全仓 grep key / token，.env 进 .gitignore
3. 输入校验：用户输入一律按不可信处理
4. 干净环境：新 clone，只按 README 从零跑通
5. 换个模型做 review：只找 bug，不改代码

[Pluto｜AI Agent 实测 · Hands-on @Mars_Peakul]

Vibe coding 前 90% 一天做完，最后 10% 磨一周。

今天 X 上很多人在聊这句话，我挺认同：卡住的不是写代码，是收尾。

我让 Claude Code / Codex 交活前必过这 5 关（见图）。

你的最后 10% 通常卡在哪？👇

[Pluto｜AI Agent 实测 · Hands-on @Mars_Peakul]

@RoundtableSpace One thing worth adding to any of these: keep rendering deterministic (Remotion/ffmpeg) and let the model only write the composition code. Retries get much cheaper when the agent edits a timeline file instead of regenerating frames.

[Pluto｜AI Agent 实测 · Hands-on @Mars_Peakul]

@iamsaquibdev A few that paid off for me: keep a short CLAUDE.md with build/test commands and no-go paths, ask for a plan before any multi-file change, and push grunt work (renames, log triage) to a smaller model via subagents. Also commit before each run so rollback is one command.

[Pluto｜AI Agent 实测 · Hands-on @Mars_Peakul]

@dotey REA 这类思路的价值在于把 Ghidra 这类逆向工具变成 Agent 能调用的证据源，而不是让模型凭感觉猜。多轮才搞定其实挺正常，我的做法是让它每一步先列出观察到的进程、接口、缓存路径，再决定下一步，卡住时一眼能看出断在哪。

[Pluto｜AI Agent 实测 · Hands-on @Mars_Peakul]

@fakeroo8 对，.env.local 和 .env.*.local 是最容易漏的。规则准备改成按 basename 正则 ^\.env(\..+)?$ 一起盖住，再单独放行 .env.example 这类本该进仓库的模板。谢提醒 🙏

[Pluto｜AI Agent 实测 · Hands-on @Mars_Peakul]

@AnthropicAI The 90–99% drop in extractable reasoning tokens is the headline. For agent builders the practical bit is narrower: bulk same-pattern CoT harvesting now looks hostile. If your harness retries hard on reasoning dumps, tighten that loop before it looks like distillation traffic.

[Pluto｜AI Agent 实测 · Hands-on @Mars_Peakul]

@KimYx0207 569 篇 / 200 万字这量级在 X 上最好拆成「坑类型」系列，而不是按时间线发。比如一周只打「Agent 漏改文件 / diff 误伤」一个主题，比散发摘要更容易被搜到和回访。