The self-distillation gameplay of Codex has become popular. An OpenAI employee personally teaches: You can make AI eliminate repetitive labor by simply copying and pasting.

A programmer from Codex has directly revealed the secret of Codex's "self-distillation"...

Just copy and paste a prompt directly into Codex.

It will actively search through your historical conversations, find out what you've been doing manually and repeatedly, and then package these workflows into reusable tools for you.

This Codex member can be said to be quite generous. After receiving feedback from netizens, he immediately iterated to a 2.0 version.

Well, what can I say... This guy is really all-round.

Greg Brockman, the President of OpenAI, also came to retweet and like it personally.

Interestingly, Brockman also retweeted a piece of trivia at the same time: Codex is open-source.

How does Codex achieve self-distillation?

Let's take a closer look at the two versions of prompts released by Vabihav (let's call him VB for short).

The first version is like this:

Look at my recent conversations and find out the repeated workflows or repeated requests.

Give me suggestions for what I've been doing manually:

1. If it's a reusable workflow, create a skill; 2. If it's a bounded role or investigation task, create a custom subagent.

Pay special attention to practical work such as CI failures, PR reviews, changelogs, document updates, release preparations, debugging, and test triage.

Only create useful things and keep it concise.

It's very short, only 9 lines.

Just copy and paste this prompt into Codex, and it will start to search through your historical conversations to find out those actions you've been doing repeatedly.

Then it will tell you which thing should be made into a skill and which thing is more suitable to be assigned to a subagent to be in charge of.

Well, this prompt is concise.

But terms like CI failures, PR reviews, and changelogs... They seem too professional, like prompts exclusive to programmers.

Sure enough, after the first version was released, a lot of feedback poured into the comment section immediately.

VB didn't idle either. He updated the second version on the same day in a hurry.

He directly expanded the data sources and the coverage scope.

The prompt also became much longer, from 9 lines to 35 lines:

The Chinese version is also posted here:

Review my work records in the past 30 days and find out the repeated manual workflows worth packaging.

Use the following data sources in order of priority: recent Codex conversations and task summaries; Codex Memories and summary records to find repeated patterns across conversations; if Chronicle is enabled, use it to discover repeated work outside of Codex (only for discovery, important details need to be confirmed in the original system); existing skills, custom agents, and automations. Prioritize reuse or expansion and don't reinvent the wheel.

Cast a wide net: coding, research, writing, planning, communication, operation, analysis, and personal affairs all count.

Only take action when all of the following conditions are met: it has occurred at least twice, or it is clearly likely to occur again and the cost of repetition is high; there is a stable input, a repeatable process, and a clear output; it can substantially improve speed, quality, or reliability; and the existing tools have not covered it. Choose the smallest applicable form: Skill, Custom subagent, Automation, Skip.

First output a candidate list, then create high-confidence projects, and finally tell me what has been created, what has been skipped, and what needs more evidence.

Compared with the first version, there are mainly three upgrades:

The data sources have been expanded from "recent conversations" to Memories and Chronicle, and Codex can see a lot more things.

The coverage scope has been expanded from coding tasks to writing, planning, communication, and operation, almost covering all your work scenarios.

(Not just for programmers. People who write articles, do operations, or do planning can all use it.)

The output has also been upgraded from "giving suggestions" to "direct creation". For high-confidence items, it will take action on its own.

The packaged results are divided into four types:

Fixed weekly repeated workflows should be made into Skills;

Investigation tasks that require a specific role to handle should be assigned to Subagents;

Regularly triggered inspections that don't require human supervision should be made into Automations;

Things that have only occurred once and have vague boundaries should be skipped directly and not packaged.

The reaction in the comment section also confirms the practicality of this prompt. Some netizens shouted on the spot:

"Make it into a plugin called /dream and pay me royalties."

VB himself also admitted in the comments that it should indeed be made into an official plugin, but he doesn't quite like the name "dream".

However, some netizens questioned that looking back at 30 days of historical records would consume a lot of tokens.?

It has to be said that tokens are really the most concerning issue for everyone...

However, VB didn't give a response. Maybe as an internal staff member, he has good support in terms of tokens.

Moreover, OpenAI has been frantically resetting Codex's rate limit recently. Friends who are interested can still give it a try boldly.

Some netizens found after trying it that half of the automatically generated skills came from "things that were done twice when the input was not stable yet".

As a result, the cost of maintaining these abstractions is higher than doing them directly again.

His question is: Should we let Codex decide whether to package or manually review?

Which new Codex features does it rely on?

Maybe you've noticed.

The Chronicle, Memory, and Subagent involved in VB's prompt are all new features that Codex has released in the past two months.

The latest one is Chronicle, which was launched on April 20th and is currently in preview.

Its function is screen capture. If you enable this function, Codex can see what you've done in the browser, Slack, and emails.

Of course, it can also see the repeated work you've done.

However, Chronicle is currently only available to ChatGPT Pro subscribers on macOS, and the memories are stored locally without encryption, which will quickly consume the rate limit and has a higher risk of prompt injection.

The Memory function was launched on April 16th along with a major update of Codex and is also a preview feature.

It can remember user preferences, correction records, and project-specific knowledge and automatically call them in future conversations.

The Subagent function was launched earlier, in March.

It is coordinated by a management Agent with multiple specialized coding Agents, and each Agent runs in an independent cloud sandbox.

VB, an OpenAI employee who loves to have fun

Finally, let's introduce Vaibhav Srivastav. He is also a very interesting person.

He is a member of the OpenAI Codex team and a former member of HuggingFace.

He used to be a machine learning developer evangelist and the head of developer experience and community at HuggingFace.

He should be one of the most active content creators in the Codex ecosystem.

On his X account, there are all imprints of him using Codex to handle daily affairs.

Just yesterday, he was still trying to use Codex to configure his Raspberry Pi to ensure that the device can be accessed remotely after connecting to the home WiFi.

On May 22nd, he claimed that "he hasn't opened an IDE for more than a month" and that using Codex alone is completely enough.

"One app to rule them all!"

He is really the most loyal fan of Codex.

In addition to this self-distillation prompt, from his tweets, Liangziwei has also discovered many interesting ways to use Codex.

For example, the /goal command of Codex: Set a "completed state" for Codex, tell it what success looks like, and then let it keep running until the goal is achieved.

There is a