OpenAI's tip for its new GPT-6.2: Tell it how you'll judge the work.
Describe the task as you would brief a colleague familiar with the project, then say what you'll be looking for when you review it. This is the core of OpenAI's guide to guiding GPT-6.2 in the codex. His example: If an icon needs to be identified in a 24px toolbar, say so.
The rest of the guide:
Adjust the effort definition to the role. Medium for a quick survey, high to trace a single suspect factor across multiple systems, especially high when multiple causes may interact.
If a mission splits into independent parts, tell the codex to "use sub-agents as needed".
For visual work, give her a reference image and ask her to check the result at a size that people will see.
Keep AGENTS.md, a codex file of project instructions, lean: only what it can't manage from the code, tests, or docs.
OpenAI offers Sol for everyday coding and writing, with fewer credits per token, and Astra for demanding work like cross-service bug tracking.