OpenAI's developer blog on September 14 advised developers to systematically review three kinds of instructions when moving to GPT-6 Astra: skill descriptions, repository-level AGENTS.md files, and task prompts themselves. The author is OpenAI's Eric Provencher, per THE DECODER, and the core judgment is that as models get more capable, the scaffolding built to walk them by the hand now gets in the way.
[1][2]Skill descriptions: short, and precise about when they apply
Skills are prompt packages stored as Markdown; their names and descriptions enter the model's context so it can pick among them. The post names two common failure modes: when descriptions are long and skills are many, Codex truncates the descriptions, so the model sees less of each; and when descriptions contradict each other or claim overly broad scope, the model loads instructions irrelevant to the task.
The post's contrast is concrete. Bad: "Create and validate Postgres schema migrations. Use when working with databases, queries, models, or persistence" — an invitation to fire the skill whenever the model touches anything database-shaped. Good: "Use when adding or changing a migration, or reviewing its rollout." The second principle is progressive disclosure: for multi-workflow skills, the root document should be a minimal router pointing to supporting docs and scripts, so the model never reads what the moment doesn't need — every extra read pushes the context closer to compaction.
One easily missed warning: repository skills constrain not only your agent but your collaborators' agents, which may run different models. Elaborate itineraries tuned for GPT-5.6 Sol or Luna may already overconstrain Astra.
[1]AGENTS.md: mandatory reading lists are context killers
The post calls out a familiar rule: "Before every edit, read architecture.md, database.md, and deployment.md." For a typo fix, that is pure waste — Astra can work out what it needs to read on its own. The right pattern is contextual signposting: architecture docs for service boundaries, database docs for schema changes, deployment docs when preparing a release.
Another practical suggestion: grant explicit permission for clearly safe operations. If local tests use throwaway data with no production access, AGENTS.md can say the agent may run tests, fix errors caused by the requested change and re-run affected tests without asking again.
[1][2]The new habit: Astra clocks out early
The post concedes a behavioral change: even without restrictions, GPT-6 Astra may stop working earlier than GPT-5.6 Sol. Worse, strict approval rules written for an era when models went rogue may now be followed with absurd literalism — you want it to keep going, and it stops, obediently.
The fix is to define "done" in the prompt: if the agent should implement, run, check results and fix errors, each step needs to be written out. Asking it to check in after the first implementation amounts to setting an earlier stopping point by hand.
[1][2]How to read the advice
The guide is useful, and worth reading on one more level. Two years ago the industry was selling prompt engineering, skill marketplaces and AGENTS.md boilerplate; now the vendor itself declares your stockpiled instructions a liability. The durable skill was never writing more instructions — it is writing termination criteria and keeping docs fresh, neither of which depends on any model generation.
Notice also the quiet shift in accountability: when the model underperforms, "your instructions are stale" is an explanation that never expires. Astra's improved judgment is, for now, OpenAI's claim about OpenAI's model. The right move for a team is to A/B its own real workflows rather than take the vendor's word wholesale — even when that word happens to be teaching you to question old words.
[1]