GPT-6 Astra calls for leaner Codex prompts

OpenAI has published a guide to updating Codex instructions for GPT-6 Astra. According to the company, scaffolding written for older coding models can degrade Astra’s performance by consuming context, triggering irrelevant skills, prompting redundant work, or causing unnecessary pauses.

Codex can load reusable skills, repository instructions, and task-specific prompts. Teams often expanded those layers to compensate for earlier models that needed detailed procedures and frequent reminders. Astra requires a review of that accumulated guidance, with shorter skill metadata, conditional repository rules, scoped permissions, and explicit completion criteria.

Stop overloading the skill router

Codex skills are Markdown instructions that may include supporting documents, resources, and scripts. Each registered skill contributes a name and description to the model’s context so the router, which selects the relevant skill, can decide when to invoke it. When too many descriptions compete for space, Codex may truncate them and route tasks using incomplete information.

OpenAI’s revised $skill-creator guidance recommends three changes:

• Trigger on concrete operations. Scope a migration skill to adding or changing migrations. Broad references to databases, queries, models, and persistence can activate it during loosely related work.
• Load detail progressively. For a skill with several workflows, keep the root document small and use it to route the model to focused guides or scripts. Codex then loads detailed instructions when the task requires them.
• Remove rigid recipes. Preserve requirements, constraints, and unusual project behavior while cutting procedures the model can infer. Excessive sequencing can block valid approaches.

Skills committed to a repository may be used by teammates running Sol, Luna, Astra, or other agents. Model-specific instructions can therefore constrain another model’s behavior, making concise and portable guidance more useful than prompts tuned to one agent.

Make repository rules earn their tokens

An AGENTS.md file supplies repository guidance on every applicable task, so stale instructions impose a recurring cost. OpenAI identifies two common candidates for removal or revision:

1. Mandatory pre-reading. Requiring architecture, database, and deployment documents before every change wastes context on small tasks. Conditional directions such as read database.md before schema changes preserve the relevant guidance.
2. Generic testing reminders. Astra usually runs tests and checks its work without repeated prompting. Keep project-specific commands, required suites, and validation constraints; remove generic instructions that merely say to test everything.

OpenAI recommends granting scoped permission for safe, reversible workflows. A repository rule can authorize local test runs against disposable fixtures without approval at each step while continuing to require confirmation for production access, destructive migrations, external side effects, or operations involving credentials.

Give Astra room, then define done

The guide describes Astra as better at judging when it can proceed safely, which changes how teams should write permission boundaries. It also reports that Astra may return after an initial implementation when a broad request leaves the stopping point unclear.

Behavior	Legacy prompt risk	Recommended update
Decision boundaries	Broad ask before doing X rules can stop safe, reversible work.	Preapprove specific local actions and reserve approval gates for destructive, costly, external, or security-sensitive operations.
Persistence	A vague task may end after the first working implementation.	Define completion to include running the code, inspecting output, executing required checks, and fixing failures.

Production controls, security requirements, cost limits, and compliance rules should remain explicit. The cleanup targets obsolete handholding and indiscriminate warnings while preserving boundaries tied to concrete risks.

Run the cleanup as an audit

Teams can apply the guidance through a focused repository review:

1. Inventory skill names, descriptions, root documents, and repository instructions loaded on common tasks.
2. Rewrite skill triggers around specific operations and remove broad topic lists.
3. Split multi-workflow skills into a small router document with focused supporting files.
4. Replace mandatory pre-reads with conditional references tied to relevant changes.
5. Keep exact test commands and project constraints while removing generic reminders Astra already follows.
6. Grant scoped approval for known-safe local actions and retain gates around concrete risks.
7. Add completion criteria to task prompts, including execution, inspection, verification, and repair.

OpenAI also suggests asking Astra to audit the repository’s existing scaffolding against the new guidance. Treat those edits like code changes: inspect the diff, verify that project-specific constraints remain intact, and test the revised instructions on representative tasks before adopting them across the team.