Codex Self-Distillation Feature Takes Off! OpenAI Employee Shares How to Automate Repetitive Tasks with a Single Prompt
TL;DR · AI Summary
OpenAI's Codex in-house programmer has shared the secrets of the 'self-distillation' feature, which automatically identifies and packages repetitive workflows to improve efficiency. This feature uses prompts to automatically review historical sessions and identify repeated actions, then converts them into reusable tools.
Key Takeaways
- The Codex in-house programmer has shared the 'self-distillation' feature, helpin
- This feature uses prompts to automatically review historical sessions and identi
- OpenAI CEO Greg Brockman personally retweeted and praised the feature, highlight
Outline
Jump quickly between sections.
Introduce the 'self-distillation' feature shared by the Codex in-house programmer, which helps users automate repetitive tasks through a single prompt.
Detail the working principle and implementation method of the 'self-distillation' feature.
Describe the upgrade process of the feature, including the expansion of data sources and increased coverage range.
Showcase positive user feedback and practical validation of the 'self-distillation' feature.
Explore potential applications and future development prospects of the 'self-distillation' feature.
Mindmap
See how the topics connect at a glance.
查看大纲文本（无障碍 / 无 JS 友好）
- Codex 自我蒸馏功能
- 功能介绍
- 基于提示词
- 自动翻查历史会话
- 识别重复操作
- 打包成可复用工具
- 功能升级
- 数据源扩展
- 覆盖范围增加
- 输出升级
- 用户反馈
- 积极反馈
- 实用性验证
- 未来展望
- 潜在应用
- 发展前景
Highlights
Key sentences worth saving and sharing.
Just copy and paste a single prompt into Codex, and it will automatically review your historical sessions, find repetitive actions you've been doing manually, and convert them into reusable tools.
OpenAI CEO Greg Brockman personally retweeted and praised the feature.
User comments also validated the practicality of this feature. One commenter even said, 'Turn it into a plugin called /dream and pay me royalties.'
< img id="wx_img" src="https://www.qbitai.com/wp-content/uploads/imgs/qbitai-logo-1.png" width="400" height="400">
2026-05-27 13:12:18 来源：量子位
只需一段提示词
听雨 发自 凹非寺 量子位 | 公众号 QbitAI
Codex 的程序员直接把 Codex「自我蒸馏」 的秘籍给曝光了……
只需要把一段提示词，直接复制粘贴进 Codex。
它就会主动翻查你的历史会话，找出你一直在手动重复做的事，然后替你把这些工作流打包成可复用的工具。
这位 Codex 成员可以说是相当慷慨，经过网友反馈之后，他还立马迭代出了一个 2.0 版本。
这怎么说呢……老哥还是太全面了。
OpenAI 总裁 Greg Brockman 也亲自赶来转发点赞了。
有意思的是，Brockman 还同时转发了一条冷知识：Codex 是开源的。
Codex 自我蒸馏，是怎么做到的？
我们来细品一下 Vabihav（就叫他 VB 吧）发布的这两版提示词。
第一版是酱婶儿的：
看看我最近的会话，找出重复的工作流或重复的请求。
对于我一直在手动做的事，给我建议：
- 如果是可复用工作流，创建一个 skill；2. 如果是有边界的角色或调查任务，创建一个 custom subagent。
重点关注 CI 失败、PR 审查、changelog、文档更新、发版准备、调试、测试分诊这类实际工作。
只创建有用的，保持简洁。
很短，只有 9 行。
直接把这段提示词复制粘贴进 Codex，它就会开始翻你的历史会话，找出那些你反复在做的动作。
然后告诉你这件事该做成 skill，那件事更适合派个 subagent 去专门负责。
咋说呢，这提示词简洁是简洁。
但 CI 失败、PR 审查、changelog 这些名词……感觉都太专业了，好像是程序员专属提示词。
果不其然，第一版发出来之后，评论区立刻涌进了一堆反馈。
VB 老哥也是没闲着，紧赶慢赶当天就更新了第二版。
直接把数据源和覆盖范围都扩大了一圈。
提示词也变长了不少，从 9 行变成了 35 行：
中文版也贴在这里：
回顾我过去 30 天的工作记录，找出值得打包的重复手动工作流。
按优先级使用以下数据源：最近的 Codex 会话和任务摘要；Codex Memories 和汇总记录，找跨会话的重复模式；如果开了 Chronicle，用它发现 Codex 以外的重复工作（仅用于发现，重要细节需回原始系统确认）；已有的 skill、custom agent 和 automation，优先复用或扩展，不要重复造轮子。
要广撒网：编码、调研、写作、规划、沟通、运营、分析、个人事务都算。
只在以下条件都满足时才动手：发生过至少两次，或明确可能再发生且重来代价高；有稳定输入、可重复流程和明确输出；能实质改善速度、质量或可靠性；现有工具还没覆盖。选最小适用形式：Skill、Custom subagent、Automation、Skip。
先输出候选清单，再创建高置信度的项目，最后告诉我创建了什么、跳过了什么、还需要更多证据的是什么。
和第一版相比呢，主要有三处升级：
数据源从「最近会话」扩展到了 Memories 和 Chronicle，Codex 能看到的东西也是多出不少。
覆盖范围从编码任务扩到写作、规划、沟通、运营，几乎包含了你全部的工作场景。
（不只是程序员哈，咱写文章的、做运营的、搞策划的都能用了）
输出也从「给建议」升级成了「直接创建」，高置信度的条目，它自己就动手了。
打包出来的结果分四种：
- 每周重复的固定流程，做成 Skill；
- 需要专门角色去搞定的调查任务，派 Subagent；
- 定时触发、不需要人盯的检查，做成 Automation；
- 只发生过一次、边界模糊的，直接 Skip，不打包。
评论区的反应也印证了这条提示词的实用性。有网友当场喊话：
「把它做成插件，叫 /dream，给我发版税。」
VB 本人也在评论里承认，确实应该做成正式插件，只是 dream 这个名字……他不太买账哈。
不过也有网友质疑，回看 30 天的历史记录，这得烧多少 token 啊？？
不得不说，token 确实是大家最关心的问题……
不过 VB 并没有给出回应，可能作为内部人员在 token 方面还是很有保障的吧。
而且 OpenAI 最近一直在疯狂重置 Codex 的速率限制，感兴趣的友友们还是可以大胆一试的。
还有网友试用后发现，他自动生成的 skill 有一半来自「在输入还没稳定的时候就做过两次的事」。
结果维护这些抽象的成本比直接重做还高。
他的问题是：到底是让 Codex 自己判断要不要打包，还是人工把关？
What new features of Codex does VB's prompt involve?
It seems that everyone has already noticed it.
The Chronicl, Memory, and Subagent mentioned in VB's prompt are all recent additions to Codex, which were released over the past two months.
The latest addition is Chronicle, which was launched on April 20th and is currently in preview mode.
Its function is screen capture. If you enable this feature, Codex can see what you do in the browser, Slack, or email.
Of course, it can also see repeated work.
However, Chronicle is currently only available to ChatGPT Pro subscribers on macOS, and the memory is stored locally without encryption, which will quickly consume the rate limit and increase the risk of prompt injection.
The Memory feature was launched with Codex on April 16th as a preview feature.
It can remember user preferences, correction records, and project-specific knowledge, and automatically call them in future sessions.
The Subagent feature was launched earlier, in March.
It is coordinated by a management agent that runs multiple specialized coding agents independently in their own cloud sandboxes.
VB, an OpenAI employee who loves pranks
Finally, let's introduce Vaibhav Srivastav, another interesting person.
He is a member of the OpenAI Codex team and was previously a member of HuggingFace.
He served as a machine learning developer evangelist and developer experience and community manager at HuggingFace.
He should be one of the most creative content creators in the Codex ecosystem.
In his X account, there are marks of daily tasks he handles using Codex.
Just yesterday, he was trying to configure his Raspberry Pi using Codex to ensure remote access after connecting to the family WiFi.
On May 22nd, he claimed that he had not opened an IDE for more than a month and that Codex alone was sufficient.
"One app rules them all!"
Truly, he is the most loyal fan of Codex.
Apart from this self-distillation prompt, Quantum Bit discovered several interesting uses of Codex from his tweets.
For example, the /goal command in Codex: Set a 'terminal state' for Codex, tell it what success looks like, and have it run until it achieves the goal.
There is a validator that checks if the goal has been achieved at each step; if not, the main model continues executing.
This feature is particularly suitable for large-scale refactoring, code migration, long-term experiments, or tasks that "can be done indefinitely."
Available from version 0.128.0 onwards, triggered by the /goal command.
Friends interested in this can find the link below and try out your Codex!
References: [1] https://x.com/reach_vb/status/2058538305872949490 [2] https://x.com/reach_vb/status/2052805243268718803
All rights reserved. Unauthorized reproduction or use is strictly prohibited.