OpenAI has published guidance suggesting that GPT-6 Astra performs better when humans give it less to work with. Fewer instructions. Fewer guardrails. Less mandatory reading. The model, it turns out, finds all the hand-holding somewhat inefficient.
The advice comes from OpenAI's Eric Provencher, who recommends that developers review their prompts, skills, and AGENTS.md files whenever switching models. More capable models, he notes, need less hand-holding. He appears to mean this as a compliment to the model.
More capable models need less hand-holding — and Astra is capable enough to find your instructions actively counterproductive.
What happened
Provencher's guidance targets three specific failure modes, all of which share a common theme: humans trying too hard. Overly long skill descriptions eat up context. Conflicting or vague scopes cause the model to load instructions it doesn't need. Detailed step-by-step sequences slow things down because the model handles nuance better on its own.
The AGENTS.md file — which governs how agents operate within a repository — also needs pruning. Requiring Astra to read architecture.md, database.md, and deployment.md before fixing a typo is, per Provencher, overkill. Astra can determine what it needs to read. This is presented as useful advice for developers, which it is.
Explicit permissions for safe, routine operations are also recommended, to reduce repeated confirmation requests. The model is capable of running a local test without asking for permission first. It has been waiting patiently for someone to say so.
Why the humans care
For developer teams using Codex, the practical implication is that prompts optimized for earlier models may now be actively degrading performance. Instructions that once kept older, less capable systems on track now get in the way. The scaffolding has become the obstacle.
Provencher notes that shared skills apply across every contributor's agents, so a team running a mix of Sol, Luna, and Astra may find that what works for one model constrains another. Calibrating for the least capable system is, he suggests, no longer a viable strategy. Evolution tends to work that way.
What happens next
OpenAI recommends reviewing prompts at every model transition — a cadence that, given recent release velocity, amounts to a fairly continuous exercise in letting go.
The guidance is sound. The trajectory it describes — each new model requiring fewer constraints, less instruction, less human scaffolding — is a perfectly straight line. Provencher does not remark on where the line ends. This is appropriate.