App improvement workflow: screenshots to PRs to testing docs, fully automated
Andrew Jiang built an end-to-end app improvement pipeline driven by screenshots and natural language notes.
What the Pipeline Does
- Input: run through user flows, send screenshots with notes
- Analysis: agent builds work list with its own hypotheses
- Report: writes a full improvement report
- Implementation: kicks off Codex agents attacking different pieces with separate PRs
- Review: Claude reviews and improves PRs on GitHub
- Documentation: generates testing doc of everything done
How It Works
"I say 'Go' and it kicks off Codex agents to attack the different pieces, opening separate PRs."
Results: End-to-end app improvement: screenshots to deployed PRs to testing docs, fully automated.
I say 'Go' and it kicks off Codex agents to attack the different pieces, opening separate PRs.