Skip prompt engineering: let Opus orchestrate cheap models like Haiku at scale
teortaxesTex · x · 2026-11-06
The author shares an Opus 5.5 session screenshot arguing that instead of learning prompt engineering, you can have a strong model orchestrate cheap ones: Opus ran 16 builds, blind code-judge workflows, and measured that a 90-word toolbox recipe cut Haiku's per-run cost from $0.24 to $0.17 with 20/20 clean builds. The takeaway: when small models are cheap enough, strong-model-driven sweeps beat hand-tuned prompts.
More from coding & agent
- Developer ships Drivr Glass: voice-first Grok cabin bot now displays cards in-car — bfrench · 2026-11-06
- Traveler says Grok bot planned and paid for his months-long Asia trip from just two context files — eptwts · 2026-11-06
- Aware lets you summon AI anywhere with @aware instead of a chat window — HaijunXia · 2026-11-06
- GitHub MCP server to replace server instructions with Agent Skills, aiming to save tokens — DanWahlin · 2026-11-06
- GitHub launches Copilot Fusion, a runtime router that auto-picks models and execution patterns — adnan_hashmi · 2026-11-06
- DIY AI Employee Factory: The 'Employee Zero' Template That Survives Crashes and Model Swaps — leebase65 · 2026-11-06