Product feedback for Windsurf

This page shows how to feed visual product reviews into Windsurf without losing the threads between findings. For other coding agents in the same playbook, see the full feedback for AI coding agents hub.

The problem with feedback in Windsurf workflows

Windsurf's Cascade flow is built for changes that span multiple files, pages, components, styles, tests. The way visual feedback usually arrives breaks that strength. A reviewer drops a screenshot in Slack, types two sentences, and moves on. By the time you sit down in Windsurf, the screenshot is one finding with no relationship to anything else. The Cascade has nothing to "span" because each finding is its own conversational fragment.

The fix is to give Windsurf one structured artifact that names every finding, links each to its source page, and provides enough context for Cascade to plan the multi-file edits in a single pass instead of one fix at a time.

The CobaltCapture workflow with Windsurf

Walk the staging build. At cobaltcapture.com, hit Capture screen, crop to the broken region, dictate the problem in plain English. Each finding lands in a single markdown document with its own H2, embedded screenshot, source URL, and commentary block.

Publish. Either copy the URL into Windsurf:

The full design review is here:
https://cobaltcapture.com/r/<slug>

Plan a Cascade across the components that surface in items 1-4.
Show me the edit list before applying.

Or download the markdown and reference it as a file in your Cascade, Windsurf's multi-file context window picks up feedback.md alongside the source files it's editing.

Why this works for Windsurf

Cascade plans best when the inputs are structured. A markdown document with explicit headings, source URLs, and bounded findings reads as a job spec, not a chat fragment. Windsurf can see "five findings, three of them in the checkout flow, two in onboarding" and plan its edit-and-test loop accordingly.

The voice dictation matters here in a way it does not for single-fix workflows. Cascade benefits from the why of each finding. "This is a recurring pattern from the design system, not a one-off," or "this fires on the wrong event because of how the new router transitions." That kind of context is the difference between a one-component patch and a real refactor. It is slow to type and natural to say out loud.

Alternatives and tradeoffs

You could feed Windsurf screenshots one by one. That works for a single bug. For a design review where the issues compound, you lose the cross-file picture.

You could write a design-review doc in Notion and paste the URL. Windsurf can fetch URLs, but Notion is not markdown and the screenshots are not embedded the way a markdown viewer renders them. The agent ends up making a second hop just to get the visuals, and sometimes fails entirely.

You could rely on Loom and your own narration. Windsurf cannot watch the video. The narration that made it useful for a teammate is dead weight for the agent.

CobaltCapture sits between "ad hoc screenshot in Slack" and "fully filed Linear tickets", exactly where most visual reviews live, and exactly the screen-capture-to-markdown pattern Windsurf's Cascade is designed to consume.