正在打开每日关注
正在打开每日关注
通过历史会话回测技能，识别真实工作流偏差并在确认后迭代。
Codex tip: take a skill or automation workflow you already wrote, throw it back at Codex, and have it run a 𝗯𝗮𝗰𝗸𝘁𝗲𝘀𝘁 against your past sessions. it rewrites the skill into how you actually work. 𝗰𝗼𝗽𝘆 𝗽𝗮𝘀𝘁𝗲 𝗽𝗿𝗼𝗺𝗽𝘁 𝗮𝘁 𝘁𝗵𝗲 𝗯𝗼𝘁𝘁𝗼𝗺. ・once a skill is done, have Codex evaluate it against the relevant sessions ・it finds the steps you always skip, the work you keep patching manually, and the order you actually run things in ・rerun this every few weeks and the skill keeps up with how your work changes the call behind it: skills should grow out of real work history. what you write from memory is the flowchart. sessions record how you actually did it and where the rework happened. session logs used to be leftovers from getting work done. now they double as an 𝗲𝘃𝗮𝗹 𝘀𝗲𝘁 for your skills. the way people write skills is changing. more builders are letting the agent read its own session history and reshape its own tools around it. 𝗽𝗿𝗼𝗺𝗽𝘁: read [skill name, workflow file, or folder path] and run a backtest against my past sessions related to it. prefer sessions from the same project and same task type. first list the samples you plan to use and why you picked them. if you can't access enough sessions, say so directly instead of guessing. compare what the skill assumes vs how i actually work. focus on: - steps i often skip, rewrite, or run more than once - work i still do manually that the skill never covers - the order i actually do things in - spots that keep causing rework, getting stuck, or needing extra clarification separate stable patterns from one-off exceptions. never rewrite a rule over a single anomaly. output a dry run first, nothing else: - keep / modify / add / remove - session evidence for each item - proposed diff and reasoning do not modify any files until i confirm. after i confirm, update the skill, run existing checks, and summarize before / after. if there's no existing way to verify, say so. finally, based on how fast new sessions accumulate, suggest a weekly or biweekly review. show me the schedule, scope, and trigger conditions first, then create the recurring task after i confirm. every scheduled review outputs a dry run only. no auto-modifying the skill without confirmation.
是什么 这是一条关于 Agent 工程与研发效能的实践建议：把已有 skill 或自动化工作流交给 Codex，用过去同项目、同任务类型的会话做回测，检查规则与真实行为的偏差，再在确认后迭代。
核心机制 它把会话日志从“完成任务后的残留记录”转为评测集。流程包括先列出样本及选择理由，再对照 skill 假设和实际执行，识别被跳过、重复、手工补做、顺序变化及反复卡住的环节；输出 keep/modify/add/remove 的 dry run，用户确认后才改文件、跑检查，并周期性复盘。
关键洞察 价值不在让 Agent 凭空自我进化，而在用真实工作轨迹校准工具契约：技能应描述稳定的操作模式、验证点和人工介入边界，而不是作者记忆中的理想流程。把回测与改动审批分离，也让优化具备可审计性。
边界与风险 会话样本可能不足、含有隐私或偶发异常；单次偏离不能直接升级为规则，必须区分稳定模式与例外。回测还可能受任务难度、项目状态和模型版本影响，需保持样本选择透明，并继续人工确认、运行现有检查。定期 review 适合只生成 dry run，避免自动修改技能造成不可控漂移。