Codex prompting

Codex Prompting

Use this skill before handing a non-trivial task to Cowork Codex through /delegate or codex_delegate.

The guidance is model-neutral and intended for current configured Codex models. Do not hard-code a model or effort level unless the user asks. Improve the prompt contract first; only escalate model or effort when the task genuinely needs it.

Core Rules

• Send one clear job per Codex run. Split unrelated research, review, implementation, and roadmap work into separate jobs.
• Prompt Codex like an operator. State the task, scope, done condition, output contract, and follow-through defaults.
• Prefer compact XML-style blocks for substantial prompts because they make routing and output expectations stable.
• Add verification rules for implementation, diagnosis, and risky fixes.
• Add grounding and citation rules for research, review, and claims that could otherwise become unsupported.
• For write-capable tasks, state the allowed workspace and tell Codex to keep changes narrow.
• For follow-ups, use resume when the new instruction continues the same Codex thread; send only the delta unless the direction changed materially.

Default Shape

Use only the blocks the task needs:

<task>
Describe the concrete job, relevant repo or research context, and expected end state.
</task>

<structured_output_contract>
Return:
1. outcome
2. evidence
3. changed files or sources inspected
4. verification performed
5. residual risks or open questions
</structured_output_contract>

<default_follow_through_policy>
Default to the most reasonable low-risk interpretation and keep going.
Only stop for missing details that materially change correctness or would cause broad or irreversible changes.
</default_follow_through_policy>

Add Blocks By Task Type

• Implementation or debugging: add completeness_contract, verification_loop, and action_scope.
• Research or recommendation: add research_mode, grounding_rules, and citation_rules.
• Review or adversarial review: prefer /review or /adversarial-review; add focus text rather than replacing the built-in review contract.
• Long-running work: add progress_updates and run in the background.

Cowork Handoff Rules

• Preserve the user's actual objective. Tighten the prompt, but do not silently change the task.
• Include the target cwd or workspace context in the task text when it is not obvious.
• If the task is inspection-only, request read-only; otherwise use the default workspace-write.
• Do not ask Codex to coordinate with Cowork's editor state. Avoid overlapping write-capable jobs on the same files.
• For multiple parallel lanes, give each Codex job a distinct scope and output contract.
• Return Codex final output verbatim unless the user asks for a summary.

References

• Prompt blocks
• Prompt recipes
• Prompt anti-patterns