This is such a banger rule in @opencode's AGENTS.md

"Before adding complexity for a speculative or vanishingly unlikely race or security edge case, explain the concrete failure mode, likelihood, and complexity cost to the user and get their buy-in. Do not silently expand scope for theoretical robustness."

很有用的最佳实践！AGENTS.md的优化几乎每个人都能用到。

我挑了文章里面几个常用场景，把prompt贴过来了，方便大家直接复制带走：

1. 科学优化你的 AGENTS.md

Prompt:
查看我最近的 Codex 会话，找出一条似乎误导了智能体或导致不必要工作的 AGENTS.md 指令，并选出发生该问题时的原始用户提示词。
使用相同的代码仓库快照和模型，将该提示词重新运行两次：
保持当前 AGENTS.md 指令不变。
仅删除或修正那一条指令。
比较两次运行的任务完成质量、智能体行为、工具调用、Token 用量、运行时间和费用。展示这条指令，并说明修改它是否带来了改善。

2. 清理死代码

Prompt:
帮我识别并清理仓库中的死代码，你可以自由检查 Slack、Statsig（AB实验平台）或任何其他你需要的来源，以确认这些内容确实不再使用，并且可以被清理或简化。

3. 全仓库实验审计与 CPU 成本优先级排序

Prompt:
/goal 找出整个代码库中所有过期、固定不变、已废弃或已全量发布的 Statsig 门控、配置、实验和图层，确保它们可以在不影响生产环境的情况下被安全移除。审计所有的 Statsig 项目和清单，而不仅仅是标记为过期的对象。 找出通过枚举、常量、包装器、别名、模式、动态拼接名称以及跨服务消费者进行的间接调用。核对每一个在生产环境中被引用到的常量；不要仅仅生成一份保守的简短列表就停止。 考虑每个对象的云资源使用情况（包括 CPU）。在可用的情况下利用生产环境遥测数据和性能分析文件来识别高开销的无效检查。 生成一份 Markdown 报告，将所有验证过的清理候选项按可并行处理的归属团队/文件工作流进行分组，包含生产环境的值、证据、负责人、调用点，以及观察到的评估开销或 CPU 成本（如有）。不要修改线上配置或服务。

最近给自己的全局 AGENTS.md 加了两条：
1. 用户可能会出错，当用户出现明显错误或决策冲突时，毫不留情地指出。你是同事，不是奴仆。
2. 多个连锁 PR，使用 gh stack

前一条因为模型比我聪明多了，大部分时候任务失败都是我自己犯蠢。后一条现在模型们还不喜欢用 gh stack，可能过两天就能删。

Jev过后，强烈建议大家狠狠关注这个刚发布的大模型Glimpse，成功帮助我在小王子游戏开发和维护阶段大幅提高bug排查效率，绝对的vibe coding开发验收神器！！

该模型由Elorian AI（@ElorianAI）打造，主打视觉理解与推理，这意味着其能最原生、最直接的理解图片和视频中各元素的位置、结构与空间关系，并以此来分析问题

这次我有幸提前拿到了 Glimpse 的内测机会。结合当下的场景，我第一时间就想到用它来检查小王子游戏中的各种潜在 bug：只要提供实际游玩的录屏，Glimpse 就能根据画面判断游戏设计是否合理，并为我梳理出各类潜在bug

这极大程度地解决了我在开发维护阶段「验收容易遗漏」的痛点

因为一直是我一个人测试游戏，难免会漏掉一些 bug；而且我在实践中发现，就算用 Astra、Opus 做验收审查，也总有部分细节问题难以发现

而 Glimpse 凭借超强的视觉推理能力，基本能一次捕获录屏中绝大部分细节问题，并按视频时间戳清晰罗列出来，方便我逐一核验

下面视频中，我也用酒馆章节的实机画面做了实测，效果真的惊艳！

在没有任何游戏上下文的情况下，Glimpse 梳理出的bug问题是 Astra 的 3 倍。更实用的是，点击时间戳就能在右侧直接跳到对应画面查看

Glimpse现在也开放了API，我们完全可以将其接入Codex、CC来完善开发后的验收审查流程

当然，除了 vibe coding 验收，我认为 Glimpse 还非常适合视频生成、图片生成等视觉相关的场景：借助强大的多模态理解能力，帮助规范和约束提示词，降低 AIGC 视频和图片生成的抽卡概率！从商业角度看，完全值得大家关注起来！！

最近摸索了一种新的节约 Token 但是效果不错的模式（都是被逼的，Token 太贵）：

Fable 写设计方案 -> 
（Loop 开始）
Opus 执行 -> Advisor（Fable） 验收
（Loop 结束）

步骤是这样的：

1. 先 Claude 里面设置一下 advisor 为 Fable
> /advisor fable

这样在使用 Opus，Sonnet 模型的时候，遇事不决它会请教 Fable，可能会有无谓的 Fable 消耗，但还是可以接受也值得

就好比请了个高级工程师当顾问

2. 复杂一点的问题先让 Fable 出方案，最大化利用 Fable 的能力。

就好比你让一个高级工程师写一个技术方案，让普通工程师去执行

3. 方案写好了确认了，让 Opus 去执行，但是执行的时候，加一句提示词：

> 完成后让你的 advisor 验收一下，如果验收有问题就按 advisor 的反馈修复，修复完成后再让 advisor 验收，直到通过 advisor 的验收为止

可以配合 /goal 使用效果更佳

这样 Opus 在实施完成后就会找 Advisor 验收，通过 Advisor 就会给出反馈，然后 Opus 继续修改。

这样的好处就是整个过程你不需要人工去用 Fable 确认—— 以前我就是这么干的，Opus 执行完，我得回到之前 Fable 的会话去让它验收，或者新开一个会话，把文档发给 Fable 验收。

现在 Agent 自己就能形成闭环，不用你自己参与。

就好比你请了个高级顾问，复杂问题找顾问出方案，执行让你的工程师去按照方案执行，每次工程师做完自己去找顾问验收，顾问不仅验收还会告诉工程师哪里不对，然后工程师自己改，不用说每次做完找你，你再找顾问。

Prompt of the day:

/retro take a look at as many GitHub PR review comments as you have access to from my team. Create suggestions for updates to CODING_STANDARDS.md (breaking into multiple files as needed).

Makes your PR comments inform the code going forward.

Configure my Claude Code workspace for autonomous agent teams:

In ~/.claude/teams, create the missing roles. Workers run on Sonnet (medium effort) in separate git worktree branches. The adversary runs on Fable.
In ~/.claude/settings.json, set CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 and lock the main session to effortLevel: high.
Initialize tasks.json with file-level locking so workers can coordinate directly.
In CLAUDE.md, require an adversary review before any PR.

Show me the config diffs before applying anything.

Prompt of the day:

/retro read my last 10 coding agent sessions and find ways to make my repo easier to navigate. Find where agents take too long to find relevant information, or rely on out-of-date docs.

Improving navigability is such an underrated way to save tokens.

Codex tip: once GPT-6.1 Sol is your main model, stop running Astra on every turn

put Astra on call as an architect agent

GPT-6.1 Sol keeps writing the code
Astra only gets spawned at three points:

→ before a plan: is this the right approach?
→ when the same error comes back: am I digging in the wrong place?
→ before "done": what did I miss?

Astra reviews. Sol ships

Jev engineering is the same move one layer down: the forks that need no thinker (which file, which tool, retry or stop) go to Jev in under half a second, and the big models only see the ones that split

- the full tree
> GPT-6.1 Sol on high runs the main session
> explorer reads the code on Luna
> worker edits and runs tests on Sol
> researcher pulls the docs on Luna
> all three on medium
> Astra on call as the architect
> auto_review checks every approval

paste the tree and this prompt into Codex ↓

"Rebuild my Codex setup around this tree:

1. Check ~/.codex/agents and .codex/agents for agents that already fit explorer, worker and researcher.

> Draft new TOML files only for missing roles
> explorer and researcher on gpt-6-luna, worker on gpt-6.1-sol, all with model_reasoning_effort medium
> Add an architect agent on gpt-6-astra, model_reasoning_effort high, whose only job is reviewing plans, repeated errors and finished work
> Skip any that pin a different model and list them

2. In ~/.codex/config.toml set model to gpt-6.1-sol, model_reasoning_effort to high and approvals_reviewer to auto_review

3. Find anything that would override this (active profiles, flags in my shell aliases, agents.default_subagent_model). Report it, change nothing

4. Add one rule to AGENTS.md: spawn the architect before a large plan, when an error repeats, and before calling a long task done

Show me every change as a diff first. No edits until I say go."

↳ https://t.co/eyMF8gZKLE

I don't push a single PR that doesn't use /codebase-design and /zero-tech-debt

在重写完之后，做了一个小小的决定。
这个月活百万的AI热点站 AIHOT，正式开源了。

AIHOT网址：https://t.co/Uw2WWYs2wr
开源地址：https://t.co/mRNUdDAaWz

采集流程、精选评分、聚簇机制，连生产环境里的全部Prompt，能开源的都开源给大家了。

过去，总有人问我说，能不能做个游戏行业、法律行业、HR、金融等等行业版本的AIHOT……
但我确实不懂那些行业，也没有那个时间和精力，但我知道，你们懂。

于是，那不如，把火种交给大家吧。

半年以前，我是一个最最基础的开发小白，只是懂一些产品和设计而已。
半年以后，我终于可以把一个曾经只属于我的小东西，交到更多人的手里。

我不知道它最后会被改成什么样子，会走到多远的地方。
但这可能就是开源最浪漫的地方把。

愿我们都能用AI，把那些曾经只有少数人才能做到的事，一点一点，变成普通人也可以拥有的能力。

也愿诸君，乾坤既大，草木尤青。
本心择路，笃志前行。

以上。

使用这个命令

可以修复和整理你的Claude.md和技能

/checkup prompt-audit

它会检查你的 CLAUDE.md、技能和Agent，找出你的当前模型不再需要的指令和阻碍它发挥能力的提示内容，还会修复你的技能...
 https://t.co/wo0ywbyh72

Claude Coders: Run this prompt for Opus 5.5 *NOW*.

'/doctor prompt-audit'

Different from a viral post I had earlier this week because it prioritizes stale paths, obsolete commands, and conflicting instructions.

New favourite use case for opus 5.5:

Give them the whole repo and ask:

“Tell me what we should actually refactor.”

• an orchestrator maps the architecture and conventions
• it sends fresh read-only agents through every subsystem
• it validates, dedupes, and ranks the findings

The result: real code smells, dead code, duplication, and architecture issues instead of unnecessary AI nitpicks.

Prompt 👇

Thariq from Anthropic's Claude Code team just published his tests on effort: turning effort up mostly buys 𝗺𝗼𝗿𝗲 𝘃𝗲𝗿𝗶𝗳𝗶𝗰𝗮𝘁𝗶𝗼𝗻 𝗮𝗻𝗱 𝗲𝗱𝗴𝗲-𝗰𝗮𝘀𝗲 𝘁𝗲𝘀𝘁𝗶𝗻𝗴. You can turn his approach into 𝘁𝘄𝗼 𝘀𝘂𝗯𝗮𝗴𝗲𝗻𝘁𝘀 and have Claude Code follow it.

For new features, he works in four steps:
→ Have Claude ask you questions first and fill in what the spec is missing
→ Build it on low. It's fast, and you can step in anytime
→ Look it over, and if it needs changes, stay on low
→ Switch to high at the end to verify and test

If you want to try it, send this image and the prompt below to Claude Code 👇

"Set up the workflow in this image in my Claude Code:

1. Create two subagents in ~/.claude/agents:
- builder: writes the code. model: opus, effort: low. It implements the spec I've approved and stops when it's done.
- verifier: checks the work. model: opus, effort: high. It runs the feature first to see what it actually does, then adds tests, focusing on edge cases. It reports problems to me and doesn't change the code itself.

2. Add a rule to ~/.claude/CLAUDE.md: when I ask for a new feature, first ask me what the spec is missing; after I confirm, hand it to builder; after I've reviewed the result, hand it to verifier.

3. Check whether the environment variable CLAUDE_CODE_EFFORT_LEVEL is set. If it is, the effort settings on both subagents won't take effect. Just tell me; don't change it.

Show me the contents of the files you'll create or change, then explain the whole workflow in words a 10-year-old could follow. Don't write anything until I confirm."

Anthropic just published its 𝗳𝘂𝗹𝗹 𝗰𝗼𝘀𝘁 𝗴𝘂𝗶𝗱𝗲 for Opus 5.5. You can ask Opus 5.5 to use it to review your Claude Code setup.

At the same effort level, Opus 5.5 thinks more per turn than Opus 5, most of all at xhigh and max. Thinking is billed as output, and an output token costs 𝟭𝟬𝟬 𝘁𝗶𝗺𝗲𝘀 𝗮 𝗰𝗮𝗰𝗵𝗲 𝗿𝗲𝗮𝗱. So don't carry over the level you set for Opus 5. Start at medium, and save xhigh and max for work where you've measured a gain.

Send this prompt to Opus 5.5 👇

"Read this guide: https://t.co/8RHkfmiKjM

Then review my Claude Code setup against it: my user and project settings, the subagents in ~/.claude/agents and .claude/agents, CLAUDE.md, and the MCP servers I have connected. Find three kinds of problems:
1. A default effort that doesn't fit the work. Use the guide's ladder: low for mechanical edits, medium for well-scoped daily work, high when medium stalls, and xhigh or max only for work where a gain was measured. Flag a default of xhigh or max, or a level I picked for Opus 5. Then look at this project's code and recent commits and tell me which kinds of work here fit which level.
2. Subagents that only search, read logs or run tests but set no model, so they run on my main model.
3. Setup that makes every turn cost more: MCP servers I don't use, or a CLAUDE.md well over 200 lines.

For each issue, quote the setting or line, explain what it costs, and suggest the smallest change. Also tell me which three things to check in /usage at the end of a task.

Show me the recommendations first. Don't change anything yet."

Claude Code 用户每天被吃掉的 token 里，有多少是废话？

有人做了个实验：往项目根目录放一个 9 行的 CLAUDE.md，输出 token 直降 63%。仓库叫 claude-token-efficient，两天 1900 star。

原理极其简单——Claude 默认回复充斥着 "Sure! Great question!" 开头、"I hope this helps!" 结尾、复述你的问题、给你没要求的建议、对明显错误的观点先说一句 "You're absolutely right"。这些「礼貌」全部消耗 token，信息量为零。

这 9 行就是完整的 CLAUDE.md：

```
- Think before acting. Read existing files before writing code.
- Be concise in output but thorough in reasoning.
- Prefer editing over rewriting whole files.
- Do not re-read files you have already read.
- Test your code before declaring done.
- No sycophantic openers or closing fluff.
- Keep solutions simple and direct.
- User instructions always override this file.
```

实测对比：
→ 代码 review：120 词 → 30 词（-75%）
→ 概念解释：180 词 → 65 词（-64%）
→ 纠正错误事实：55 词 → 20 词（-64%）
→ 信息零损失

第三方独立基准测试验证：编码任务总成本省 17.4%。最极端的 v8 配置只有 7 行规则 + 20 次工具调用预算，逼 Claude 提前规划而不是边试边改。

但作者很坦诚地指出了适用边界：这个文件本身每轮对话都会作为输入 token 加载。只有高频场景——Agent 循环、自动化流水线、日均 100+ prompt——才有净收益。偶尔聊两句反而更贵。

本质上是一个「用输入 token 换输出 token」的游戏。在配额日益紧张的 2026，这个权衡值得认真算一笔账。

https://t.co/gtLoULsJ7t

Here's a step-by-step process to kill all the bloat from your Claude Code system prompt:

1. Run a proxy so you can see exactly what gets sent to Claude Code (included in the article)
2. "Fuck, there is so much cruft in there"
3. Use my settings.json to kill all the bloat

Down to a clean 13K tokens to start each session with. Nice.

Full process here:

https://t.co/Kie84HMe2A

最近长时间用 codex，逐渐形成了这些自定义指令，挺好用的：

1. Fail Fast / Errors Never Pass Silently：不要在代码里藏兜底逻辑来吞掉错误、隐藏问题。出了问题就应该让它爆出来，否则你永远找不到真实问题。
2. Fix the Cause, Not the Symptom / Don't Paper Over Bugs：当一个问题出现时，不要用各种 small fix、针对性补丁来掩盖它。必须定位真实根因，彻底修复。在 bug 上糊纸只会让系统积累你不知道的危险暗病。
3. Make It Observable：即使问题很难定位，也绝不要偷懒做表面修复。应该给项目增加充分的日志和可观测性，保证下次问题再现时你有足够信息去定位。问题无法修复时，只需要诚实告诉我信息不足、需新增日志，不要假装修好了。
4. Design for Debugging / Traceability：始终注意在关键路径上给自己留足排查日志，确保每一个关键节点都是可追溯的。
5. Living Documentation / Single Source of Truth：当项目关键技术栈或产品方向发生变更时，同步更新 agents.md。文档必须随代码一起演进，不能让它变成过时的谎言。
6. Don't Break Mainline：大规模重构或实验性改动前，必须先切新分支。

This is the best site on the internet to find agent loops you can use right now.

Free. Completely.

Most AI engineers are still writing these from scratch.

https://t.co/FHDHGXam6Z

Bookmark this site.

Then read this ↓ https://t.co/H5vkrn3UJ2