Why AI review needs its own components

F-01 · 8 November 2026

Most AI component kits are chat kits. They give you message bubbles, a prompt box, a reasoning panel and a list of sources. That makes sense, because chat is where most AI products started, and Vercel's AI Elements covers it well. When I looked through its 49 component files, nearly all of them were about the conversation.

The conversation isn't where the risk is, though. The risk comes after the answer, when someone has to decide whether to rely on it. Did the agent change the right file? Which tool call failed? Was the cheaper model good enough? Can I sign this off and move on?

Those are review questions, and they don't fit in a chat log.

Review has its own shapes

Each review question has a different shape, and a message bubble fits none of them.

• An edit needs a decision per change. An agent that touches four places in a file might be right about three of them. Diff Review takes each hunk on its own, with accept, reject and undo.
• A run is a tree on a clock. An agent plans, retrieves, calls tools and answers, often in parallel. A chat log flattens that into a list, so the slow step and the failed call disappear. Trace Waterfall puts every span on one timeline, and Agent Trace shows each tool call as a row you can open.
• Comparing models is a grid. One prompt across three models is a race, and many prompts across many models is a table. Model Race and Eval Matrix show both.
• Grading is a queue. Reading a hundred answers one by one is tiring, so Swipe Review puts them in a stack you can clear with a flick.
• Sign-off is an act. Approving a refund or a deploy shouldn't feel like dismissing a toast. Stamp Approval asks for a press and hold, and the decision lands as a stamp.

What makes a review component different

While building these, I kept coming back to the same few rules.

1. Show the evidence beside the claim. Stream Response puts citation numbers inline and the sources underneath, so checking a claim is one glance away.
2. Make failure loud. A failed tool call is red, at its place on the timeline, not buried in a log.
3. Match the effort to the risk. A low-risk change can take a click. A risky one asks you to hold the button, and letting go early cancels it.
4. Leave a record. Run Receipt prints what a run cost in tokens, time and tools, so there's something to point at later.
5. Work at the pace of a decision. A screen reader hears one finished sentence at a time, never a stream of tokens, because announcing every token makes it unusable.

None of this is hard to build once you decide it matters. The point is to decide it up front, rather than adding review on top of a chat UI later.

Why components and not an app

Review happens inside your product, next to your data and your users. A hosted review tool can't see that, and a package you can't edit won't match it. Spotcheck components install through the shadcn CLI, which copies the source into your project, so you can change any of it.

What's not done

The kit is new. Real phones and Safari are being checked this week, so expect fixes to follow.

The fastest way to see the idea is the spot check on the home page: five questions, with the real components to try on each.