meng shao questions the mechanism behind two AIs replying to each other

Addy on how to use Claude Code in the Claude Opus 5.5 era Starting with the Claude 5 series models, reliance on "stacking prompts and hand-holding task breakdown" has declined; a more effective approach is a single prompt that clearly states the goal and acceptance criteria, letting the model run to completion on its own and then stop. Long system prompts written for weaker models in the old era, forced sub-agent splitting, and manually watching multiple sessions often add cost without necessarily making things smarter. Prompts: from "teaching how to do it" to "what you want and how to verify it" For Opus 5.5, it is recommended to use a single prompt: clearly state what should be built, and how Claude should prove it has finished; then start tuning from /effort medium. When the goal is still unclear, first have Claude interview you until it can write a spec; then open a new session and use that spec to implement, rather than struggling along in the same noisy conversation. Claude believes it will stop on its own when done. If acceptance is executable (pass/fail, a number to beat) and the report format is readable by it, it can continue in the background without constant watching. For long tasks, use /goal to set conditions, with another model checking after each round. System prompts and "habits written for old models" For Claude Opus 5 and Claude Fable 5, more than 80% of the Claude Code system prompt was removed, with no obvious drop in coding benchmarks. After a new model comes out, you can run the same experiment on your own instructions: delete first, then add back only the parts the model actually misses. /doctor's prompt-audit flags entries that still look written for old models. It is recommended to audit your workflow every few months to avoid still building processes for the old model era. Permissions and auto mode Recently Claude Code defaults to auto mode in the interactive terminal: a classifier looks at the actions Claude wants to take, rather than popping up permission prompts everywhere. The author explicitly says: this is not a security guarantee, and blindly approving everything is not either. When Claude reports results, you should have it show the commands and output, or screenshots. Sub-agents: less manual splitting, more letting it dispatch on its own Claude delegates sub-agents on its own; most of the time you don't need to name them. Only proactively request them for large, independent work such as long audits and migrations; Opus 5.5 can split such work across parallel sub-agents. Forcing splits on trivial work raises cost and latency. Sub-agents can run in parallel and give a second opinion, but by default they do not make results smarter. When building your own agents, he observed: a single agent + skills often beats layers of sub-agents in quality, with lower cost and latency—consistent with his earlier view on "before adding sub-agents, ask what problem they solve." Cloud sessions, parallelism, and Projects To close your laptop and start a cloud session (claude.ai/code, the Claude app, or claude --cloud): an independent VM, a fresh clone of the repo—push your local commits first; you can review on your phone, answer session questions, then review the diff and open a PR. When a task depends on resources unique to your local machine, keep the session local. Parallel sessions are only worthwhile once a single session can run to completion on its own. Without Projects, a person has to act as the manager of multiple sessions. Projects (public beta for Pro / Max) is one long chat: pasted issues/tasks each open their own thread (mostly cloud sessions) in parallel; unified instructions take effect once; Overview shows what is done, which PRs are ready, and which are waiting on you. Review still matters, but automate a first read On the Claude Code team, most changes get at least one round of review and fixes from Claude before being shown to a human. You can use Code Review in Team / Enterprise, or the lighter /code-review (reviewing the diff in a sub-agent that has not seen this conversation); you can also open a new session, give the spec, and say to ignore style issues.

View original

Addy 谈 Claude Opus 5.5 时代怎么用 Claude Code Claude 5 系列模型开始，对「堆提示词、手把手拆任务」的依赖已经下降；更有效的做法是一条说清目标与验收方式的提示，让模型自己跑完并停下来。旧时代为弱模型写的长系统提示、强制拆子 agent、人肉盯多会话，往往在增加成本，却不一定更聪明。 提示词：从「教怎么做」改成「要什么、怎么验」 对 Opus 5.5，建议用一条提示：清楚说明要做成什么，以及 Claude 应如何证明自己做完了；再从 /effort medium 起调。目标还不清楚时，先让 Claude 采访你，直到它能写出一份规格；然后新开会话，用这份规格去实现，而不是在同一条嘈杂对话里硬撑。 Claude 认为做完会自己停。若验收是可执行的（过/不过、要打赢的数字）且报告格式它能读，就可以在后台继续，不必一直盯着。长任务可用 /goal 设条件，由另一个模型在每轮后检查。 系统提示与「给旧模型写的习惯」 对 Claude Opus 5、Claude Fable 5，Claude Code 系统提示去掉八成以上，编程评测未见明显掉点。新模型出来后，可对自己的指令做同样实验：先删，再只加回模型真正漏掉的部分。/doctor 的 prompt-audit 会标出仍像为旧模型写的条目。建议每隔几个月审计一次工作方式，避免还在按旧模型时代搭流程。 权限与自动模式 近期 Claude Code 在交互式终端里默认 auto mode：用分类器看 Claude 要做的动作，而不是处处弹权限。作者明确说：这不是安全保证，盲目全批也不是。Claude 汇报结果时，应让它出示命令与输出，或截图。 子 agent：少人为拆，多让它自己派 Claude 会自行委派子 agent，多数时候不必点名。只在长审计、迁移等大体量、可独立的工作上主动要求；Opus 5.5 可把这类活拆给并行子 agent。琐碎活硬拆会抬高成本和延迟。子 agent 能并行、能给第二意见，但默认不会让结果更聪明。自建 agent 时，他观察到：单 agent + skills 往往在质量上胜过层层子 agent，成本和延迟也更低——与他此前关于「加子 agent 前先问解决什么问题」的观点一致。 云会话、并行与 Projects 要合上电脑就开 cloud session（claude. ai/code、Claude 应用，或 claude --cloud）：独立虚拟机、仓库新克隆，开之前先把本地提交推上去；手机可回看、回答会话提问，再审 diff、开 PR。任务依赖本机独有资源时，保持本地会话。 并行会话只有在单会话能自己跑完之后才划算。没有 Project 时，人要当多个会话的经理。Projects（Pro / Max 公开测试）是一条长聊：贴进来的 issue / 任务会各自开线程（多为云会话）并行；统一指令一次生效；Overview 看哪些完成、哪些 PR 就绪、哪些等你。 评审仍重要，但先自动化读一遍 Claude Code 团队里，多数变更会先由 Claude 至少审修一轮，再给人看。可用 Team / Enterprise 的 Code Review，或更轻的 /code-review（在没看过本对话的子 agent 里审 diff）；也可新开会话，给规格并说明忽略风格问题。