OpenAI Publishes a Leaner Prompting Playbook for Codex Agents

OpenAI has published guidance for adapting Codex repositories to GPT-6 Astra, telling developers to cut back long skill descriptions, blanket document-reading rules and rigid approval steps. Its central argument is that scaffolding built for earlier coding models can consume useful context, trigger irrelevant work and cause Astra to pause before a task is actually finished.

The advice is not to remove instructions wholesale. OpenAI says teams should review skills, AGENTS.md repository files and task prompts when switching models, then make the remaining rules more specific. The proposed replacement for generic hand-holding is a system that tells the agent which constraints apply to a particular job and exactly what completion requires.

Keep the routing layer small

Skills are Markdown prompts that can include resources and scripts. Their names and descriptions are placed in the model’s context so Codex can select a relevant skill. OpenAI warns that too many skills can force those descriptions to be truncated, leaving the system with incomplete information for that choice.

• Describe a concrete operation, rather than a broad subject area, so a skill activates only when it is relevant.
• For a skill covering several workflows, use a short entry document that points to narrower guides or scripts when needed.
• Keep unusual project requirements and constraints, while removing detailed procedures the model can infer itself.

A repository guide should point, not pre-load

AGENTS.md files pose a different problem because they apply repository guidance across work. OpenAI recommends replacing mandatory pre-reading with conditional directions: consult architecture material for service-boundary work, database material for schema changes and deployment notes when shipping. A small typo fix should not require a full project tour.

That logic matches an earlier OpenAI engineering account of its own agent-first development work. The company said a giant instruction file crowds out the immediate task, code and relevant documentation. Its alternative was to keep a short AGENTS.md as a map to a structured, repository-based knowledge system rather than treat it as an encyclopedia.

Freedom for routine work, gates for real risk

OpenAI also draws a line between low-risk local work and actions with larger consequences. Teams can explicitly permit safe, reversible operations such as running local tests on throwaway data, fixing related errors and rerunning affected tests. But the guidance retains approval gates for production access, destructive migrations, external side effects and credential-related operations.

The tradeoff is especially clear in OpenAI’s description of Astra’s behavior. The company says inherited restrictions can be interpreted too literally, preventing the agent from continuing through work that a team considers safe. Yet removing friction does not answer a separate problem: Astra may also stop earlier than GPT-5.6 Sol when a broad prompt fails to specify the full workflow.

Define “done” before asking for autonomy

OpenAI’s answer is to replace a vague request with an explicit end state. If a task requires an implementation, the prompt should also say to run the code, inspect the output, execute required checks and fix failures. By contrast, an instruction to check in after an initial implementation establishes an earlier stopping point.

That makes the new playbook less a recipe for giving agents unrestricted latitude than a cleanup exercise in operational clarity. OpenAI suggests having Astra audit existing skills and repository instructions, then reviewing the changes and testing revised instructions on representative tasks. The unresolved practical question is whether teams can distinguish obsolete caution from the project-specific rules that still prevent costly mistakes.