GPT-6.2 Astra 难用?AGENTS.md 汇报契约让它变顺手
很多人吐槽 GPT-6.2 Astra 难用,其实根本不是它智商不够,是太不会跟人类沟通了。只要在 AGENTS.md 里加上一份“汇报契约”,这头性能猛兽立马变得极其顺手🤯
- 顶级智商配了个拉胯的“注意力界面”:海外开发者 Dan McAteer 一针见血指出,Astra 的原始推理能力绝对是第一梯队,但每次干完活把中间试错、死胡同和命令输出全倒给你,根本不考虑切回窗口的人类没空看废话。
- 写进 AGENTS.md 的五大强制汇报结构:要求最终回复必须按固定顺序输出——① 一两句话说清最终结论;② 需要我拍板或审批的事项(附推荐选项);③ 无需追问就能直接上手的下一步;④ 真正改了什么(砍掉试错过程);⑤ 证据与置信度清单(区分实测、推断和未验证项,必须附文件路径)。
- 模型越强越得立规矩:评论区不少重度用户直呼这是“Game Changer”,有老哥把规则浓缩成四句话同样好使。有意思的是,连作者自己在回复里都忍不住承认了一句“G4 Neon 统治一切”。
你在用 Astra 或其他 Agent 写代码时,会特意在 AGENTS.md 里限制它的回复格式吗?
直接把原帖里可直接复制进 AGENTS.md 的“汇报契约”模板放这了,用 Codex 或其他 Agent 干活的兄弟可以直接粘进去试效果:
How to report
Your final message is read by someone switching in from other work. They will read the first lines and maybe nothing else, and they cannot see your commands or their output.
Write the final message in this order, with these headings:
1. Outcome. One or two sentences: what is now true, or what you found and recommend.
2. What I need from you. Each decision or approval that is the reader's to make, numbered, with your recommendation for each. Say "nothing" if there is none.
3. What to do first. The next steps in order, numbered, each one something the reader could start on without asking a follow-up.
4. What changed or what was learned. Only the detail the reader needs to act. Leave out the path you took, dead ends and retries.
5. How sure we are. Three short lists: what you checked and against which file or page; what you are inferring or proposing rather than reading; what nobody has tested. Include any source you could not open and any command that was blocked.
Rules:
- State something as fact only if you read it or ran it in this session. Otherwise mark it as an inference.
- Give the file path or URL behind each claim the reader would act on.
- Plain words and full sentences. Shorter is better, as long as the reader could act without asking a follow-up.
本文由 AI 翻译自英文原帖,技术名词保留英文。
查看 X 原帖