Codex Self-Distillation prompt goes viral inside OpenAI, turning chat history into reusable Skills and Subagents
A viral OpenAI prompt called Codex Self-Distillation turns your chat history into reusable Skills, Subagents, and Automations.
What it actually is
An OpenAI engineer published a prompt that asks Codex to scan your recent chat history, find repeated patterns, and package them as **Skills**, **Subagents**, or **Automations** you can reuse. OpenAI president Greg Brockman amplified it. Within 48 hours it had been iterated from a 9-line script into a 35-line instruction that handles email, planning, and reporting in addition to code.
The technique has been nicknamed **Codex Self-Distillation**. It is not a model update. It is a prompt that turns Codex into a meta-tool that builds its own tools by watching what you actually do.
Why this matters more than another benchmark
Most AI productivity gains today are stuck in the same loop: a person writes a prompt, gets an output, paste-edits, and moves on. The output is disposable. Self-Distillation breaks that loop. Codex looks at the last hundred turns and asks, what did this user do more than twice. Then it builds the artifact that automates it.
The trick maps roughly to four output types:
- **Skills**: fixed, repeatable processes worth packaging as a callable tool
- **Subagents**: tasks that need investigation and dynamic decisions
- **Automations**: scheduled or event-triggered jobs
- **Skip**: things not worth packaging
That fourth bucket is what makes it usable. Most automation tools fail because they try to automate everything. This one knows when to walk away.
How OpenAI staff are actually using it
Internal anecdotes describe Codex acting as a personal chief of staff. Investor-relations dashboards. Slack summarizers. Deadline trackers. One team reportedly ran an experiment to ship a real product with zero human-written code, where the bottleneck shifted entirely to high-level design and review.
That last point is the real story. If self-distillation works at scale, the scarce resource at a software company stops being engineers who write code and starts being engineers who can review and direct agents.
What to do with it
Try the prompt against your own Codex history this week. The first run usually surfaces three or four repeated patterns you did not realize were patterns. Package the obvious ones and ignore the rest. The point is not to automate your job. The point is to spend the next quarter doing the work that did not fit into your week before.
Source: OpenAI