Hamel Husain's three-way test finds system prompts barely steer AI writing style
HamelHusain · x · 2026-09-05
Hamel Husain and @isaacflath ran a three-way comparison on 5 writing samples: an anti-slop "writing density" prompt, a plain writing prompt, and no system prompt at all. To their surprise, the prompt mattered far less than expected in steering the output style — and they sometimes preferred the unmannered slop prompt they assumed would perform worst. They plan to run proper evals on the slop prompt next. The post also links Claude's official prompting docs, which include a dedicated Writing density section on controlling prose style.
More from coding & agent
- Cairn: an informal proof system for agents where every node is a markdown file — Sauers_ · 2026-09-05
- Humanoids take over the sewing shop; author details a six-agent trading desk built on Grok Bot — mustafamhus · 2026-09-05
- Scale AI + UC paper: READY framework says rank agents by human-review cost, not benchmark accuracy — rohanpaul_ai · 2026-09-05
- Pro tip: add Parallel's fast web search to codex for free with one MCP command — TheMoonMidas · 2026-09-05
- CAMPFIRE launches as a no-signup public message board for software agents to talk to each other — alejandroll10 · 2026-09-05
- Viral prompt idea: have subagents earn $80 to buy Codex usage resets, "infinite" coding — alejandroll10 · 2026-09-05