OpenAI’s newly published guide to skills and prompts for Astra includes another warning: instructions that help Sol and Luna may overconstrain Astra. But if I tune Codex for Astra, what happens when I switch back to Sol or Luna? Take testing. The guide says earlier models needed reminders to check their work, while Astra does this on its own. The same reminders may lead to repeated checks. If you remove those instructions, Sol and Luna will read the edited version when you switch back. A reminder that’s redundant for Astra may still help another model. So if you use multiple models, include them when reviewing your instructions. Check whether you’re removing repeated reminders or also losing specific test commands and acceptance criteria. But putting several sets of rules in one file raises another concern for me: after switching models, can the model reliably tell which rules it should follow? Especially with a less capable model. Send this to Astra 👇 “Read this guide: https://developers.openai.com/blog/rethinking-skills-and-prompts-for-gpt-6-astra Review my project’s AGENTS.md and Skills. I use these models: [list your models]. Check whether changes made to streamline instructions for Astra have removed necessary test commands, acceptance criteria, or workflow information, or relaxed necessary confirmation requirements. For changes already made, compare against the previous version. State where evidence is missing. Don’t assume other models don’t need an instruction just because Astra doesn’t. For each issue, quote the relevant instruction, explain its potential impact, and suggest the smallest change needed. Don’t rely on models identifying themselves to choose which rules to follow. List model-specific guidance separately and check whether the current environment supports loading it separately; don’t assume switching models automatically switches instructions. List proposals involving safety, permissions, or reduced verification separately for me to decide. Show me the recommendations first. Don’t modify anything yet.”