A breakdown of OpenAI's three-page Astra prompting and config guide
A read of OpenAI's official Astra guide: the model's problem is stopping too early rather than running off, so old rules should be removed, with 14 rules and config settings included.
OpenAi just showed a 3-page manual on prompting and configuring GPT-6 Astra most Astra guides are wrong on day one. they teach you to control a model that runs too far. Astra has the opposite problem: it reads your instructions, takes them seriously, and stops early. so the job flips - you’re not writing guardrails anymore, you’re removing them what’s inside: → the prompts that make Astra finish instead of pausing, taken straight from OpenAI’s own model guide → the AGENTS.md lines you wrote for GPT-5.6 Sol that now sabotage Astra, with the exact replacements → a completion contract it can’t skip, enforced by a Codex Stop hook instead of a paragraph → the config.toml keys for approvals, subagents and context, with the literals that actually parse → the stops that aren’t yours to fix, and how to tell them from a bad prompt three things in it most people get wrong: the “1M context window” is 258K inside a subscription, and no key raises it Astra on low already beats GPT-5.6 Sol on high, so you’re probably overpaying for effort the model’s own “done” is not evidence, give it a verifier that doesn’t trust the report it’s short on purpose. the setup for this model is mostly deletion the 3-page manual is the summary. the full breakdown - all 14 rules, every prompt, the code and the config - is in the article below. read it before you touch another AGENTS.md, then bookmark it
This post quotes another post: x.com/adiix_official/status/2097014889990545608