Codex adds personalized prompt predictions

OpenAI is rolling out composer predictions for Codex, giving the coding agent a way to suggest a user’s next message. The beta feature uses the current conversation and the user’s phrasing style to generate a follow-up prompt in the composer. It is currently available to ChatGPT Pro subscribers.

OpenAI’s developer team says composer predictions became one of its most popular features during internal testing. The appeal is straightforward: coding sessions often produce predictable follow-up requests, so a relevant suggestion can reduce repeated typing while leaving the next instruction under the user’s control.

A familiar request takes shape

A recent GitHub request, issue #42587, proposed similar behavior after Codex completes a turn. The author suggested context-aware ghost text or a composer chip, keyboard shortcuts for accepting or dismissing it, and a safeguard preventing suggestions from running automatically.

The released feature addresses the request’s core idea, although OpenAI has not confirmed whether it uses the proposed interface or keyboard controls. The company also has not described how users can disable predictions or manage the personalization data behind them.

Where predictions save keystrokes

Agentic coding sessions commonly cycle through implementation, testing, debugging, review, and commits. After Codex changes a codebase, the next instruction may be easy to anticipate: run the tests, inspect failures, review the diff, or open a pull request. Composer predictions can place that likely instruction in the input field before the user types it.

Personalization should make those suggestions resemble each user’s normal commands. Someone who writes “run tests” may receive terse predictions, while a developer who usually specifies test targets, constraints, and expected output may see more detailed prompts.

Rollout details and open questions

• Availability: Beta rollout for ChatGPT Pro subscribers.
• Subscription: Included with the $200-per-month Pro plan.
• Location: The Codex composer in the ChatGPT desktop app.
• Other clients: OpenAI has not announced support for the Codex CLI or IDE extension.
• Controls: Acceptance shortcuts, dismissal behavior, and settings have not been documented.

Best fit: iterative coding work

Long sessions provide the clearest use case because Codex can infer the next step from an established sequence of tasks. Predictions may help when triaging test failures, applying a refactor across files, reviewing generated changes, or preparing a pull request.

One-shot questions and exploratory conversations offer less history for prediction, so suggestions may be less reliable in those cases. The feature’s practical value will depend on how accurately it anticipates intent, how quickly users can accept or dismiss a suggestion, and whether OpenAI brings the same workflow to its CLI and editor integrations.