Product feedback for v0

This page is for designers and engineers iterating on a Vercel v0 build who need to hand the chat structured visual feedback. The same pattern applies across the feedback for AI coding agents hub, but v0's chat-first surface has its own quirks worth naming directly.

The problem with feedback in v0 workflows

v0 outputs React components. You prompt, it generates, you preview, you reply with corrections. The hard part is that reply. Subtle problems, the hero copy is breaking on a 1280px viewport, the secondary button uses a near-miss shade from the design system, the form spacing doubled since the last revision, the icon is one tier too large next to its label, are easy to see on the live preview and brutal to describe in a chat sentence.

So you end up writing "make the button smaller and fix the spacing," v0 picks one of those, lands a different shade of wrong, and the iteration drifts. After five rounds you're further from the design than you started, and you've burned credits getting there. The chat surface optimizes for short messages, but the feedback you actually need to give is structured, multi-finding, and visual. Without a way to attach all of that to one short message, every iteration loses ground.

The CobaltCapture workflow with v0

Open your v0 preview in one tab. Open cobaltcapture.com in another and hit Capture screen. Pick the v0 preview window. Drag a box around the broken hero. Hit Dictate and talk through it out loud. "The eyebrow copy should sit closer to the headline, the headline is using the wrong font weight, and the gradient behind the CTA is the wrong direction, push it to the bottom-right." Repeat for the navbar. Repeat for the form. Hit Publish.

You get a public URL like cobaltcapture.com/r/<slug> with every screenshot cropped, every finding captioned, and your dictated reasoning written out under each frame. The source URL of the v0 preview travels with each item, so v0 always knows which revision you're commenting on.

Back in v0's chat, use the Add prompt to send the next iteration. Paste the URL, name the order, and ask v0 to confirm each change before moving on. The chat stays one message long. The structured payload travels through it. If you'd rather hand v0 raw text, click the markdown export and paste the body directly, the embedded image links render inline in the chat the same way.

Example prompt

I've documented three issues with the latest revision here:
https://cobaltcapture.com/r/<slug>

Apply them in order. Confirm what you changed and which component
it landed on before moving to the next.

Why this works for v0

Three reasons. First, v0's chat-based interface accepts pasted URLs and follows them, the screenshots referenced inline in the markdown get processed as visual context for the next generation, not dropped as opaque attachments the model has to hunt for. Second, v0 is good at translating natural-language design intent into JSX, and dictation captures the kind of "why" a typed-from-thumb sentence skips. "Use the secondary style here because this is a low-stakes confirmation" produces a different patch than "fix the button." Third, the Add prompt pattern rewards short messages with rich context behind them, which is exactly what a CobaltCapture URL is. One line in the chat, a dozen findings behind the link, every finding with its own cropped frame and dictated reasoning. The credit you save on retries pays for the workflow many times over.

Alternatives and tradeoffs

You could paste raw screenshots straight into the v0 chat. That loses the source URL each frame came from and loses the dictated reasoning that explains why the change is needed. v0 ends up guessing intent from pixels.

You could record a Loom and link it. v0 can't watch video. The narration that would have made the recording useful is invisible to the agent, this is the same handoff problem that makes structured screen capture to markdown the right format for any agent-in-the-loop workflow.

You could keep describing issues in chat. That works for one finding. It collapses at five, where you skip details to keep the message short, and each skipped detail becomes a future iteration that produces a worse preview. CobaltCapture is what keeps the iteration count low.