Deny

@hellodeny

Founder and product designer at Craftwork.

Bangkok

Deny
@hellodeny
Jul 14
Replit has described a loop on their agent: production failures → clusters → hypotheses → patches → evidence → ship. you can't afford their infra. but you already have the production traces: your own transcripts. retro runs the same loop on your laptop. for free. tonight.
00:00
Deny
@hellodeny
Jul 13
$ /retro all scanned 433 sessions. last 45 days. 655 failed tool calls 224 corrections typed by hand 74 × "you're right" 40 × Esc mid-action 8 sessions abandoned all recorded to disk. never read. you never read yours either. that changes today. retro.craftwork.design
00:00
1
Deny
@hellodeny
Jul 9
Your agent keeps making the same mistakes because nobody reads its failures. I built retro. ✦ reads your local session transcripts ✦ finds what you correct again and again ✦ writes the CLAUDE.md rules that stop it No API keys. 100% local. MIT. retro.craftwork.design
00:00
1
Deny
@hellodeny
Jul 9
There is 1.3 GB of evidence sitting on my macbook that my AI coding agent keeps making the same mistakes. 1,850 session logs on disk. Every failure recorded. I never read a single one. Tomorrow I ship the thing that reads them.
00:00
Deny
@hellodeny
May 6
POV: you're trying to balance vibecoding, running 2 startups, posting Instagram carousels, X posts, Dribbble, Figma designing, team managing, new startup ideas, gym, visiting parents, kids caring, and having a personal life
GIF
1