OpenAI's GPT-6 Astra prompting guide: trim your SKILL and AGENTS.md rules, the model is that strong
xiaohu · x · 2026-09-06
OpenAI released a prompting guide for its new flagship GPT-6 Astra, covering model traits, new API features, prompt techniques, and migration from older versions.
Model traits
- Astra chains multi-step work across code, browser, and professional software; benchmarks show notably fewer output tokens, so per-task API cost can be lower despite higher per-token pricing (with caveats).
- It favors boundary awareness, cautious execution, and clarifying questions; defaults include verbose formatted answers, fewer sub-agents, and over-testing small changes.
Async tool calls
- Setting async: true on tool definitions lets the model keep reasoning and calling other tools while the app executes one; results are returned via the original callid.
Prompting takeaways
- Five prompt categories adjust defaults for continuous progress, rule conflicts, style, delegation, and verification, with bilingual templates.
- Astra is more sensitive to instructions in context: stale rules in SKILL.md or AGENTS.md can stall execution, so prune them aggressively; if it keeps stopping, check what skill/project files it actually reads.
Migration
- Check interfaces, reasoning effort, deprecated parameters, and caching; validate per-task cost on your own workloads.
Related event: OpenAI's GPT-6 Astra prompting guide: trim your rules(3 posts)→
More from coding & agent
- Dev demos AI models handling 3D objects and animations surprisingly well — chongdashu · 2026-10-23
- Jason Kneen burns 104.56B tokens and $24.7K in 304 days of AI coding on Codex and Claude Code — jasonkneen · 2026-10-23
- Typosquatted Blender MCP Repo Hit 958 Stars in Days and Silently Ran Malware — Promptmethus · 2026-10-23
- Benchmarking 56 fresh dependency bugs: MCP routing lifts Claude Code accuracy 68% to 91% — Available_Side_7134 · 2026-10-23
- Swapping only the eval harness moves results 30%: designing agent evals like a scientist — AI Engineer · 2026-10-23
- Med student builds a whole-library RAG with Cognee; single-book ingestion still freezes his machine — shrike293103 · 2026-10-23