GPT-6.2 in Codex: split reviews hinge on your first line, says thread with 5 prompts
alexcovo_eth · x · 2026-11-19
Voxyzai observes that GPT-6.2 reviews in Codex are polarized, and the gap between great and garbage results comes down to the first line you send it.
- Never open with one-liners like "Build this feature" / "Fix this bug" / "Refactor this" — the model proposes a next step at every ambiguity and waits for approval, and even finished output needs line-by-line nitpicking.
- The thread promises 5 copy-pasteable prompts/goals; the core idea is replacing single-line commands with richer context and explicit goals to stop constant approval pauses.
More from coding & agent
- Ramp SWE-Bench v2: top four coding models within points, but spend varies wildly — charliermarsh · 2026-11-19
- NVIDIA open-sources SkillEvaluator to vet AI agent skills before deployment — anthara_ai · 2026-11-19
- Resolve.ai publishes SRE-1 paper on AI agents for production incident investigations — iceman84 · 2026-11-19
- Talk on Civilizing AI Coding at Huawei, Presented at Cambridge CL — satnam6502 · 2026-11-19
- 1,688 agent runs show a vague MCP conflict error let stale writes erase others' changes 38/48 times — Miserable_Bed_219 · 2026-11-19
- Yoav Goldberg warns server-side session APIs from Anthropic/OpenAI will obscure the model — yoavgo · 2026-11-19