What they did
The poster, who runs a startup called Mana that gives LLMs primitives for building apps, has been benchmarking vibe-coded apps for a year. This time he gave Opus 5.5 one instruction, to build something Figma-level with feature parity so a user could not tell the difference, and let it run on his platform with very little further guidance. He says the same approach then produced clones of Google Meet, Docs, Telegram and Linear from single prompts, and shared an editable link to the Figma result.
How it went
He reports about three hours and eight prompts for the editor, and that as a casual Figma user he cannot see a difference. Commenters were split. One asked how 8,200 lines could cover Figma; the poster said to try it. A former Figma engineer who led the variables workstream said parity for casual use is believable but power-user features and edge cases are where these clones fall apart, and that the collaborative platform cannot be vibe-coded. Another commenter who tried a similar prompt gave up after 45 minutes and asked for a standard way to compare prompts, output and tokens.
Worth knowing
The poster is the founder of the harness the agent ran on, and the thread treats the post partly as an advertisement. Nobody reproduced the result independently in the thread, and the exact prompt, CLAUDE.md and settings were requested but not posted. Read it as an existence proof for casual-use parity, not a benchmark.