OpenAI's three-page Astra manual says to delete old prompts
The author says Astra's problem is stopping too early rather than running too far, and recommends removing GPT-5.6-era guardrails from AGENTS.md, defining "done" clearly, and verifying with a Stop hook.
OPENAI JUST SHOWED A 3-PAGE MANUAL FOR GPT-6 ASTRA Most Astra guides are already wrong on day one. They teach you how to control a model that runs too far. Astra has the opposite problem. It reads your instructions, takes them seriously, and often stops earlier than you expected. What actually matters: > Remove old GPT-5.6 Sol guardrails from AGENTS.md > Tell Astra clearly what “done” actually means > Give it permission to make reasonable assumptions > Use a Stop hook to verify completion > Keep approvals only where they’re actually needed > Separate bad prompting from system-level stops > Don’t trust the model’s own “done” as evidence This model doesn’t need more prompting. It usually needs less. The biggest mistake is carrying your old setup forward and assuming more instructions = better performance. The 3-page manual is the short version. The full breakdown has all 14 rules, prompts, config examples and the exact replacements. Read it before you touch another AGENTS.md.
This post quotes another post: x.com/w1nklerr/status/2096964317107077567