Humanizer skills invent facts and drop commitments; his open-source rewrite wins 48% blind test
PawelHuryn · x · 2026-10-09
PawelHuryn tested humanizer skills for workplace writing and found serious flaws: a plain "rewrite this to sound human" prompt told a customer their £45 refund had already gone through; the "undetectable" Emulate-1 dropped an auto-renew cancellation deadline; the 55K-star blader/humanizer apologized for a deposit deduction, which can read as admitting fault. He iterated for weeks and open-sourced work-humanizer (MIT), designed to keep commitments intact and ask instead of inventing. In a blind test on 12 work texts with 4 AI judges from 4 labs, work-humanizer won 48% (18 meaning changes) vs blader/humanizer 20% (33), naive prompts 4%/1%, and Emulate-1 0% (120 changes). Bonus finding: Pangram cares less about wording than reasoning structure—feeding his own writing as notes made 4/4 texts score 100% human.
More from coding & agent
- Runway MCP + Claude Opus 5.5 Demo Wows Users: "I'd Pay $120 for This" — tlakomy · 2026-10-09
- Memory Journal: A Portable .txt Format That Transfers AI Context and Skills — NoobSolid26 · 2026-10-09
- Bluerails Launches Hotel-Booking MCP Server Covering 280K Properties in DACH — bookwithbluerails · 2026-10-09
- Voyager launches as 'Codex for creative work', driving Blender, Resolve and After Effects — rohanpaul_ai · 2026-10-09
- PartyKit founder shuts down free hosted platform 2.5 years after Cloudflare acquisition — threepointone · 2026-10-09
- world-intel-mcp: open-source MCP server with 132 tools for real-time global intelligence — tom_doerr · 2026-10-09