I’m interested 😊. What are you building?

Exactly, the screenshot gate was the part I learned the hard way 😅. Appreciate you digging in 🙏

Both, and there's a hard cap so it can't loop Checks run first: learned house rules, a design detector, and screenshot checks. If the screenshots can't be trusted, it doesn't get reviewed at all. Then a fresh reviewer looks at the real screenshots and makes one call: ship, fix, rebuild or recapture. It gets 2 fix rounds by default. If it's still not a ship after that, a human decides.

loom doesn't score, impeccable's finish reviewer does, as a fresh agent with no shared history. It gets the request, the chosen direction's contract, the screenshots + their notes, and detector output, then returns a disposition: ship, fix, rebuild or recapture. No numeric rubric. It judges the real captures against the contract. Close-enough calls get 2 fix rounds, then you decide.

Free and MIT. One-line install is in the README. Then ask your agent: "use loom to redesign my landing page" github.com/evidence-codes/lo… Built on @pbakaus's impeccable 🙏

Every mistake a review catches becomes a house rule, plus an automatic check where code can detect it. So the next project starts stricter than the last.

Screenshots lie. Stretched heroes, fake phone widths, animations caught half-faded. loom captures what visitors actually see. It found that Margin, a concept page I built, overlaps its chapters for anyone with reduced motion on.

The reviewer sent it back twice before it passed. Along the way it caught: – card grids hiding inside the new design – labels stacked above headings – a screenshot frame cutting through UI – a mobile overlay that never rendered All things I'd have shipped.

loom runs a process instead of a prompt. It reads the project, offers genuinely different directions, builds the one you pick, then has an independent reviewer judge real screenshots.

The reviewer sent it back twice before it passed. Along the way it caught: – card grids hiding inside the new design – labels stacked above headings – a screenshot frame cutting through UI – a mobile overlay that never rendered All things I'd have shipped.

loom runs a process instead of a prompt. It reads the project, offers genuinely different directions, builds the one you pick, then has an independent reviewer judge real screenshots.

Proof: my own portfolio. Before, it had the classic AI look: dark gradient, card grid, one glowing accent. loom turned it into an indigo resist-dyed cloth with a live WebGL hero. evidence.codetribehq.com

AI-built sites all look the same. I built an open-source workflow that fixes that. loom gives your coding agent distinct design directions, honest screenshots, and a reviewer that sends the work back until it passes. Works in Claude Code, Cursor, Codex and Antigravity 🧵

The reviewer sent it back twice before it passed. Along the way it caught: – card grids hiding inside the new design – labels stacked above headings – a screenshot frame cutting through UI – a mobile overlay that never rendered All things I'd have shipped.

loom runs a process instead of a prompt. It reads the project, offers genuinely different directions, builds the one you pick, then has an independent reviewer judge real screenshots.

LFG 🚀🚀

LFG 🚀🚀🚀

Applying for this one, and their application asked for a quick video on something I learned recently. Made this one about noticing Claude Code @claudeai uses words like "cogitated" and "discombobulated" in its reasoning, turns out that's the model describing its own confusion-to-understanding process as it works through code. Enjoyed making it enough that I want to turn "something I learned recently" into an actual series. Feels like good practice for growth work too, taking something dense and making it click for someone without the context. Video below.

evidence-codes

Very much interested