OpenAI Codex Team Releases Guide for Reviewing Skills and AGENTS.md for GPT-6 Astra

The OpenAI Codex team (Eric Provencher) published a blog post recommending the removal of unnecessary instructions from repository and user settings ahead of the transition to GPT-6 Astra. Improved model performance means that excessive scaffolding from the past (such as encouraging tests or using prohibition phrases) now causes context waste and instruction conflicts.

Skill Management and Context Budget

Skills consist of Markdown-based prompts and reference materials, adhering to the Agent Skills open standard. The initial skill list is limited to a maximum of 2% of the model's context window; if exceeded, descriptions must be shortened or skills excluded. Skill descriptions should place core use cases and trigger words at the beginning to improve matching accuracy.

AGENTS.md and Instruction Design Principles

Since AGENTS.md is always applied during repository work, conditional instructions are recommended. Unconditional instructions to read documentation slow down performance. GPT-6 Astra has excellent alignment, so strong prohibition phrases like 'Never' or 'Ask first' can cause tasks to stop prematurely. For safe workflows, it is effective to explicitly state 'proceed without approval' along with the rationale.

Checkpoints When Transitioning Model Generations

Both OpenAI and Anthropic emphasize that harness instructions must be regularly reviewed when swapping model generations. Anthropic demonstrated that removing more than 80% of the system prompt in Claude 5 resulted in no performance loss. For GPT-6 Astra, clearly defining 'completion' before starting a task and removing unnecessary review request phrases are necessary to resolve Persistence issues.