The review-loop guides · a straight answer
How do I get human feedback on an AI-generated web page?
You built it with an agent in twenty minutes. Now you need a human to tell you what's wrong with it — fast, and without setting up an account for them.
Short answer: share the page as a live link and let people pin comments directly onto it — no login, no Figma, no screenshots emailed back and forth. Your reviewer taps the exact spot that's off and types a note; you (or your AI agent) get those notes back as a structured list and act on them. Tools built for this — like twocents — let an agent share the page and read the feedback back automatically, so the whole loop stays inside the session that built the page.
That's the version that actually works. But it's worth seeing why the obvious alternatives don't, because most people reach for one of them first and lose an afternoon to it.
The options, honestly
| Approach | The catch | Good when |
|---|---|---|
| Screenshot + email/DM | Feedback arrives as "the button, second one down, looks weird" — you spend more time locating notes than fixing them. | One tiny change, one person. |
| Figma | Your page is real HTML, not a Figma file. Importing it loses the interactivity, and your reviewer needs an account and the know-how. | You already live in Figma. |
| A pull request | Only works if your reviewer reads code. Non-technical eyes — the ones who catch "this copy is confusing" — can't use it. | Reviewer is another dev. |
| Pin-on-the-page link | Needs a tool that hosts the live page and captures notes — but nothing to install for the reviewer, and notes come anchored to the exact element. | Anyone, on a phone, in one tap. |
The failure they share is disembodied feedback: a note that isn't attached to the pixel it's about. Half of review effort is just re-locating what the comment refers to. Pinning kills that step.
Try it — tap anywhere on the mock page to drop a pin, the way a reviewer would:
Trusted by teams everywhere
The loop that keeps it inside the build
The reason to use a purpose-built tool isn't the pinning — it's that the feedback comes back to the agent that built the page, structured enough to act on. Concretely:
- Share. Your agent calls the MCP tool with the page's HTML and gets a review link.
- Collect. You send the link to two or three people. They open it on their phones, tap the spots that are off, and type notes — no login, nothing to install.
- Apply. The agent reads the notes back (each anchored to its element), shows you each one, and edits only the ones you approve.
You stay the filter — the agent automates the noticing and the doing, never the deciding. That distinction matters: three strangers' opinions aren't a spec, and a good loop applies your judgment to their notes, not theirs to your page.
Why you can't just review it yourself
You go blind to your own page after about ten minutes on it. Whether the headline reads as confident or generic, whether a stranger knows what to click first — that's precisely the judgment you can't make about something you just made. One outside tap is worth an hour of your own squinting. That's the whole reason the feedback step belongs in the build, not after it.