Back to Podcast Digest
Every7m

How I Built an AI Software Factory With Fable 5

TL;DR

  • Overnight merges are already real in his workflow: He described the "magic moment" as starting a run at night, letting it merge automatically if CI stayed green, and waking up to Brandon praising the slicker design.

  • Fable 5 felt like a genuine step change after one week: He said it is his favorite model ever because things he previously wanted to work suddenly started working, especially for shipping features fast.

  • Slack feedback becomes structured engineering work twice a day: A scheduled Claude Code routine reads the v2 alpha channel morning and evening, classifies messages, checks what is already solved, downloads recordings, and creates a PR with YAML and markdown tracking files.

  • RiffRaff captures richer bug reports than plain video: His open source React wrapper records clicks, voice, network requests, and errors into a shareable file, giving the agent much better debugging context than a normal screen recording.

  • Batching feedback is the key review trick: Instead of reviewing 17 separate pull requests, he combines the items into one PR branch in Cursor, asks Claude to fix what it can and leave notes where human judgment is needed, then reviews the changes in one pass.

  • The system gets better by remembering mistakes: He credited the compound engineering setup with helping the agent avoid repeating the same errors when new rounds of feedback come in.

The Breakdown

He kicked off a batch of 17 Slack-reported bugs before bed, told the system to merge if CI passed, and woke up to a better-looking product and a teammate saying, "whoa." Using Fable 5, Claude, Cursor, and his own open source feedback recorder, he turned app feedback into an overnight AI software factory.

Was This Useful?

Share