OpenAI tells devs: Lean prompts, fewer guardrails for GPT-6 Astra

OpenAI has recently published detailed prompting tips for GPT-6 Astra, and these latest recommendations from Eric Provencher on skills and project instructions build on that earlier guidance. The advice focuses on two areas where developers often accumulate cruft: skill descriptions and AGENTS.md rules. Skills, which are stored as Markdown files, work best when scoped to specific workflows, but too many skills force Codex to truncate descriptions and lose the information it needs to choose correctly. Conflicting or overly broad descriptions can also load unnecessary instructions. Similarly, mandatory reading of multiple documents before every change wastes context, and the fix is to point selectively to architecture, database, or deployment docs depending on the task.

Another theme is that restrictions designed for earlier models may backfire. Provencher notes that if earlier models went rogue and you locked things down with strict approval rules, it's time to revisit those when switching to Astra. The model's better judgment means it can interpret old restrictions so literally that it stops even when you want it to keep going. Known, safe workflows should be explicitly allowed. Even without restrictions, Astra may stop earlier than GPT-5.6 Sol, so defining upfront what "done" means becomes important.

The stakes here hinge on how teams manage shared skills that apply to every contributor's agents. What works for Sol or Luna might already be too restrictive for someone running Astra, so teams face the challenge of aligning instructions across different models. The outcome depends on how carefully developers review and adapt their prompts and rules when switching models, rather than assuming one set of instructions fits all.