Your agent just built something good. A dashboard, a landing page, an interactive report — generated in a single conversation and rendering perfectly in front of you. Now you need three other people to look at it and tell you whether it ships. That hand-off is where most AI workflows fall apart.
This is a step-by-step workflow for sharing an AI-generated app for feedback without zip files, screenshots, or standing up hosting you don't need.
Why the obvious options are slow
The instinct is to screenshot the app and paste it into Slack, or record a Loom and narrate a walkthrough. Both detach the feedback from the thing it's about. A reviewer types "the second button is confusing" into a thread, and now you're matching prose to pixels by hand. Worse, by the time the comment arrives, your agent has already produced two newer versions, and nobody is sure which one was reviewed.
The other instinct is to deploy it — push it to staging, wire up a host, send a URL. For a prototype that changes every prompt, that's a lot of infrastructure for a review that should take a minute to set up. If sharing takes longer than reviewing, review happens too late to matter.
The workflow: one link, four moves
The faster path treats the share link as the workspace, not a delivery mechanism. Here is the loop, the same one described in detail on the collaborative review for AI-generated apps page.
1. Publish from the tool that built it
Whatever generated the app — Claude Code, Cursor, Lovable, v0 — publish its HTML straight to a private share link. With Capsule, your agent does this over MCP with a single instruction: "publish this to Capsule." There is no export step, no build pipeline, and no "can you send me the file again." You get a live, browser-based link immediately. If the audience is external, password-protect or time-limit it.
2. Send the link, not the artifact
Reviewers open the link in a browser. They don't need an account, they don't need your tool, and they're not looking at a stale copy pulled from a chat transcript. They see the working app exactly as you do.
3. Collect feedback on the page itself
Instead of describing where a problem is, reviewers point at it. Comments attach to the element they're about, in the state the reviewer saw it. The button that's mislabeled, the chart that's wrong, the paragraph that reads badly — each comment carries its context with it. Feedback stops being an archaeology exercise.
4. Iterate without changing the link
Here's the part that keeps the loop tight: the feedback is readable by your AI agent. Back in your tool, ask it to check the comments. It lists the open ones, proposes a fix for each, and waits. You decide which fixes to apply. Approved changes are pushed as a new version on the same link. Reviewers see what changed; you never send a new URL.
What "good" looks like
A sharing workflow worth adopting passes a few simple tests:
- Zero-setup sharing. Publishing should take under a minute, or review slips to the end of the project.
- Feedback in context. A comment should anchor to the element it concerns.
- One canonical link. Reviewers should never wonder if they're on the current version.
- Agent-readable feedback. If your AI can't read the comments, a human re-types them into the prompt — that's the broken loop again with extra steps.
- Human approval. AI proposes; a person decides what ships.
If you're using a specific tool, the mechanics are the same but the publish step differs. We've written it up for Claude artifacts, Cursor projects, and Lovable apps.
Try it on something real
The fastest way to feel the difference is to publish something you've already built. Install Capsule for your agent — it's a single setup message — and put your next AI-generated page on a review link before the conversation that generated it ends. The review stops being the slow part.