Codex Self-Distillation Goes Viral! OpenAI Employee Teaches: Copy-Paste to Let AI Eliminate Repetitive Work
OpenAI employee Vaibhav Srivastav shares a prompt that makes Codex automate your repetitive tasks — and it's spreading fast.
Article intelligence
Key points
- OpenAI employee Vaibhav published a prompt that makes Codex scan history for repetitive workflows and package them into tools
- Two versions: v1 focuses on coding, v2 expands data sources to Memories and Chronicle, covering many domains
- Relies on Codex features Chronicle, Memory, and Subagent for automation
- Vaibhav, a Codex team member, hasn't opened an IDE in over a month
Why it matters
This matters because openAI employee Vaibhav published a prompt that makes Codex scan history for repetitive workflows and package them into tools.
Technical impact
May affect model selection, inference cost, product capability, and evaluation benchmarks.
An OpenAI employee has published a 'self-distillation' trick for Codex that is quickly going viral across social media. Vaibhav Srivastav, a member of the Codex team, shared a simple prompt: copy and paste it into Codex, and the AI will automatically review your past conversations, identify manual repetitive tasks, and package them into reusable tools such as Skills, Subagents, or Automations.
The first version of the prompt was only nine lines long, focusing on technical tasks like CI failures, PR reviews, changelogs, and debugging. It instructed Codex to look at recent sessions and suggest which workflows should be turned into skills or custom subagents. After receiving user feedback, Vaibhav released an expanded version the same day. The second version spans 35 lines and broadens the data sources to include Codex Memories and the Chronicle feature, which captures screen activity on macOS. It also extends the scope from coding to writing, planning, communication, operations, and personal tasks. Notably, the new version doesn't just suggest — it directly creates high-confidence tools.
This self-distillation technique relies on three recently introduced Codex features. Chronicle, released in preview on April 20, captures your screen to detect repetitive work across browsers, Slack, and email — but it's limited to macOS ChatGPT Pro subscribers and consumes rate limits quickly. Memory, added on April 16 as a preview, remembers user preferences and project knowledge across sessions. Subagent, fully released in March, enables a managing agent to coordinate multiple specialized coding agents in independent cloud sandboxes.
Vaibhav, previously a machine learning developer advocate at Hugging Face, is one of the most active content creators in the Codex ecosystem. On his X account, he constantly shares his Codex-powered workflow. He recently claimed he hasn't opened an IDE for over a month, relying solely on Codex. He also demonstrated other tricks, such as the /goal command, which makes Codex autonomously work toward a defined completion state.
The community response has been enthusiastic but with caveats. Some users pointed out that scanning 30 days of history would burn a significant number of tokens, and others noted that automatically generated skills sometimes stem from tasks that weren't stable enough to warrant automation. Vaibhav acknowledged the feedback and suggested future improvements. OpenAI's president Greg Brockman also retweeted the prompt, adding that Codex is open source.