Prompting regime is back: scaffolding and context now separate vastly different results from same models
Afinetheorem · x · 2026-11-15
Kelsey Tuoc observes we may be (briefly?) back in a regime where prompting skill plus stored tools, context and examples produce vastly different results from the same models. Afinetheorem agrees, citing an AI-made Beethoven movie with none of the "look at camera, say 5 words" slop: with the right scaffold, context and human intervention, design, writing, video and coding all improve dramatically. Single-prompt chatbot use has fundamental limits — workflow determines quality.
More from coding & agent
- MarkItDown becomes Microsoft's most-starred GitHub repo with 196k stars, overtaking VS Code — SaleemaAmershi · 2026-11-15
- LMMs Lab open-sources The Harness: one unified framework for computer, robot, and game agents — liuziwei7 · 2026-11-15
- shadcn authors designed UI libraries by inspecting model weights, making agents better — cramforce · 2026-11-15
- Google launches Gemini Decisions API with Gemini Decision 1, zero output-token billing — sebkrier · 2026-11-15
- Claude Code advisor: Fable 5.2 adds 1.7 points for Opus 5.5 at 2.1x cost — and can hurt — socialwithaayan · 2026-11-15
- Lessons from building 5 deterministic MCP servers for finance/data tasks — Ltx_23 · 2026-11-15