In the AI field, Codex's self-distillation approach has recently attracted widespread attention. A programmer from OpenAI, Vaibhav (referred to as VB), shared his techniques, which allow Codex to eliminate repetitive tasks and improve efficiency with just simple prompts. VB revealed that by inputting a copied and pasted instruction into Codex, it will automatically scan your history chat, identify the tasks you often perform manually, and package them into reusable tools.

After receiving user feedback, VB quickly released a 2.0 version of the prompt, further enhancing its functionality. OpenAI CEO Greg Brockman also supported this feature and shared this little-known fact on social media: Codex is actually open-source.

So, how does this self-distillation work? The first version of VB's prompt was concise and aimed at identifying repeated requests in the user's recent conversation and suggesting whether to create a reusable workflow or a custom agent. Although the first version was short, its terminology was mostly specific to programmers, leading to many feedbacks, so VB expanded it.

The second version of the prompt increased from 9 lines to 35 lines, with a significant expansion of data sources and coverage. From analyzing only coding tasks, it now covers various work scenarios such as writing, planning, communication, and operations. This version of instructions not only provides suggestions but can also directly generate high-confidence projects, divided into four types: Skill for fixed processes, Subagent for investigative tasks, Automation for scheduled checks, and Skip for un-packaged tasks.

However, despite the enthusiastic response, concerns about historical records consuming tokens still exist. For this issue, VB did not give a clear response but encouraged users to try boldly. At the same time, OpenAI recently reduced the rate limits for Codex, allowing more people to experience this feature.

Finally, VB is not only a technical expert but also a loyal user of Codex. He shared his daily use of Codex on social media, even stating that he hasn't opened an IDE for over a month and relies entirely on Codex to complete his work. Clearly, his love for Codex is evident.

Key Points:

🌟 **Codex Self-Distillation Feature**: With a simple prompt, Codex can identify repetitive tasks and automatically generate reusable tools.

🚀 **Version Upgrade**: The second version of the prompt significantly expands data sources and coverage, supporting multiple work scenarios.

🛠️ **User Feedback**: Although the response has been enthusiastic, concerns about token consumption still need attention, and users are encouraged to try boldly.