shadcn spots Claude doing 'retroactive edits' — striking through its earlier wrong answer
shadcn · x · 2026-10-17
shadcn, creator of shadcn/ui, shared a novel Claude behavior: after he questioned a suspicious answer and asked Claude to verify, Claude came back, apologized for the mistake, returned to its previous answer, struck through the incorrect paragraph, and added an "[Edit: ...]" annotation.
He says it's the first time he's seen this "retroactive edits" behavior and asks whether other agent harnesses do something similar — a new form of models actively maintaining consistency across their earlier outputs.
More from Fun
- Wei Dai's X account reportedly compromised, DMing crypto scam messages — NathanpmYoung · 2026-10-17
- User Dresses AI Companion Muse in New Halloween Outfits Every Day — ChrisUniverse · 2026-10-17
- Dating in SF: "I wouldn't change anything about your markdown file" — vhanagwal · 2026-10-17
- Startup culture 2024: working hard isn't enough, you must post about it — vhanagwal · 2026-10-17
- Creator builds "Claudia x Roadster" film with Claude, MCP, Midjourney and BytePlus — elonmusk · 2026-10-17
- Gary Bernhardt mocks "Claude can do literally anything" hype with 'plz go outside' — teropa · 2026-10-17