OpenAI, GPT-6 Astra는 간결한 프롬프트와 완화된 가드레일 권장

GPT-6 Astra needs leaner prompts and fewer guardrails, OpenAI recommends Matthias Bastian View the LinkedIn Profile of Matthias Bastian Sep 12, 2026 GPT-Image-2 prompted by THE DECODER Ask about this article… Search Overly long skill descriptions, blanket reading requirements, and rigid approval rules can get in GPT-6 Astra's way, according to OpenAI. The company recommends that developers tie instructions more tightly to specific tasks and define more clearly when the job is done. Instructions that have piled up over time can eat up context or cause GPT-6 Astra to stop work too early, writes OpenAI's Eric Provencher. He recommends reviewing skills, AGENTS.md , and task prompts whenever switching models. More capable models need less hand-holding, according to Provencher. That lines up with earlier advice from OpenAI on model transitions . Vague skill descriptions lead to wrong picks Skills are prompts stored as Markdown files that can include resources and scripts. Provencher says they work best for specific workflows or applications. Their names and descriptions go into the model's context so Codex can pick the right skill for a given task. Ad Too many skills force Codex to truncate descriptions , stripping out information it needs to choose correctly. Conflicting descriptions or overly broad scope can also load instructions the model doesn't need, according to the blog post. Ad Provencher recommends keeping scope descriptions short and precise. A skill for Postgres schema migrations should only fire when creating or modifying a migration or when checking its rollout. If a skill covers multiple workflows, its main document should briefly point to the right supplementary docs and scripts, Provencher writes. That way the model only reads what it actually needs, since every extra read eats context and pushes it closer to summarization. Detailed step-by-step sequences can also slow down newer models, since they handle nuance and ambiguity better on their own, according to Provencher. But shared skills apply to every contributor's agents, so teams need to be careful. What works for Sol or Luna might already be too restrictive for someone running Astra. Ad Mandatory reading before every change wastes context The rules in AGENTS.md that govern repository work need regular reviews too, Provencher says. Requiring the model to read multiple documents or a full project overview before every change is overkill for a typo fix. Astra can figure out what it needs on its own. Instead of forcing it to read architecture.md , database.md , and deployment.md every time, Provencher suggests pointing to those docs selectively. Architecture info when working on service boundaries. Database docs when changing schemas. Deployment notes when shipping. The docs also need to stay current, he adds. Ad Explicit permissions can also cut down on repeated confirmation requests for safe operations, according to the post. For local tests using throwaway data with no production access, AGENTS.md can explicitly allow the agent to run tests, fix errors caused by the requested change, and re-run affected tests without asking again. Ad Astra needs a clear goal, not a checklist If earlier models went rogue and you locked things down with strict approval rules, Provencher says it's time to revisit those when switching to Astra. OpenAI credits the model with better judgment, but it can also interpret old restrictions so literally that it stops even when you want it to keep going. Known, safe workflows should be explicitly allowed, according to the post. Even without restrictions, Astra may stop earlier than GPT-5.6 Sol, Provencher notes. He recommends defining upfront what "done" means. If the agent should implement something, run it, check the results, and fix errors, all of that needs to be in the prompt. A requirement to check in after the first implementation sets an earlier stopping point. OpenAI recently published detailed prompting tips for GPT-6 Astra , and these skill and project instruction recommendations build on that guidance. AI News Without the Hype – Curated by Humans Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section. Subscribe now Source: OpenAI