Generating a website is no longer the hard part. An agent can produce a polished, working page from a paragraph of intent in less time than it takes to write the brief. The hard part now sits one step later: getting the right people to look at the result, react to it, and agree on what ships. That step — review — has barely changed, and it's quietly become the bottleneck.
The mismatch between generation speed and review speed
Generation collapsed from days to seconds. Review didn't. It still runs on the same improvised stack it did before AI showed up: screenshots in chat, Loom walkthroughs, screen-shares, "can you send me the file." The result is a workflow where the fast part is fast and the slow part swallows everything you saved.
The mismatch isn't a tooling oversight. It's structural. Most review tools were built for one of two artifacts, and an AI-generated website is neither.
Why existing tools don't fit
Code review evaluates source. Pull-request tools are excellent at diffs, but the people who need to sign off on a generated website — a product manager, a client, a founder — don't care about the markup. They care about the rendered page: how it reads, how it clicks, whether it's right. Reviewing the diff tells them nothing they can judge.
Staging environments evaluate deployed products. They work, but they assume infrastructure: a host, a pipeline, a deploy step. For an artifact that changes with every prompt, standing up a deploy for each iteration is absurd overhead. And even when you do, the feedback still ends up somewhere else — a thread, a doc, a meeting — detached from the page.
So teams fall back to screenshots and chat. And that's where three specific failures show up.
The three failures of the improvised stack
Feedback detaches from its context
A reviewer marks up a PNG or types "the hero CTA is weak" into Slack. The click, the state, the exact element — the context that makes the feedback actionable — is gone before it arrives. You spend more time reconstructing what they meant than fixing it.
Nobody agrees which version was reviewed
The page changed three times while the feedback was in transit. Someone reviewed a version that no longer exists. Decisions scatter across threads and calls, and there's no single record of which feedback was actually addressed.
The loop doesn't close back to the agent
This is the one that's unique to AI. The whole promise is that the agent can act on feedback instantly. But if the feedback lives in a Slack thread, a human has to read it, interpret it, and re-type it into the prompt. That manual relay is the broken loop — the same friction you had before, now with an extra translation step.
What closing the gap actually requires
Reviewing AI-generated websites well needs three properties the improvised stack can't provide. We make the full argument on the collaborative review for AI-generated apps page, but in short:
- The review surface is the live page. Feedback attaches to the element it's about, in the state the reviewer saw — not to a screenshot or a paraphrase.
- Iteration is part of the loop. Comments flow back to the agent that generated the site; the fixed version lands on the same link reviewers already have.
- Publication is a decision. A human approves a specific version, and that approval is recorded. AI proposes; a person signs off.
Put those together and review stops being a detour. The link is the workspace: the page, the comments on it, the fixes the agent proposes, and the version you approve all live in one place.
This isn't a tooling preference — it's the missing layer
Capsule isn't an AI tool and isn't a website builder. It's the collaborative layer between the tools that generate sites and the people who decide they're ready. The generation got fast; the review needs to catch up, and that means putting feedback back on the page and back in front of the agent.
If you build with a specific tool, the review mechanics are the same — only the publish step differs. We've documented reviewing Claude artifacts, Cursor projects, and Lovable apps. Or install Capsule for your agent and run your next review on the page itself.