Product feedback for Zed AI

This page is for developers using Zed who want visual product feedback the assistant panel can actually act on. It walks the workflow and the prompt pattern. For the same playbook applied to other tools, see the feedback for AI coding agents hub.

The problem with feedback in Zed workflows

Zed's AI assistant is fast and it accepts pasted context, which means the bottleneck isn't the model, it's the quality of what you paste in. Most visual feedback arrives in lossy forms. A teammate drops a screenshot in a Slack thread with two sentences underneath. A designer DMs a paragraph about a layout issue with no image attached. A QA pass produces a wall of bullets in a Linear comment with no source URL.

When you sit down in Zed and try to brief the assistant, you end up retyping. The dictated nuance from a Loom is gone. The URL where the issue lives is gone. You paste the screenshot, type a sentence summarizing what someone said better in voice, and the assistant does its best with the typed sentence. The patch lands close but not quite right, because the input was already a downgrade.

The CobaltCapture workflow with Zed

Open the staging build you want feedback on. Open cobaltcapture.com in a new tab and click Capture screen. Pick the staging window. Drag a box around the broken component, then hit Dictate and talk through the problem. "The modal dismisses on outside click, but the form has unsaved state, so the user loses input every time they accidentally tap the overlay." Repeat for each finding. Hit Publish.

You get a public URL like cobaltcapture.com/r/<slug> and a one-click markdown export. In Zed, open the assistant panel (cmd-?), then paste the URL:

The design review for the checkout flow is documented here:
https://cobaltcapture.com/r/abc12345

Please plan edits across the components that surface in items 1-4.
List the file changes before applying.

Or download the export as feedback.md, drop it into your project, and reference it directly in the assistant panel. Zed's multi-file editing flow picks up feedback.md alongside the source files it's changing, so the assistant sees the findings and the code together in one context window.

Example prompt

@feedback.md

Work through items 1, 2, and 4. Skip item 3, that's intentional.
For each item, show me the file diff before moving on.

Why this works for Zed

Zed's AI assistant is at its best when you paste structured markdown context. A document with one H2 per finding gives the assistant a clear job spec instead of a chat fragment. The assistant reads "five findings, three in checkout, two in onboarding" and plans accordingly.

Embedded screenshot URLs render inline when the assistant fetches the page, so the visual context arrives in the same paste as the text. No second hop, no attachment-juggling.

Dictation is the part that's hard to recreate any other way. When you talk through a finding. "This fires on the wrong event because the new router transitions asynchronously", you capture the kind of nuance Zed's typed-from-thumb input box tends to miss. That context is what turns a one-line fix into a real refactor when a refactor is what the problem needs.

The assistant also handles model swaps cleanly. Whether your Zed session is wired to Claude, OpenAI, or another provider, the markdown input is the same artifact. Switching models doesn't change the prompt, which means the workflow survives every settings tweak you make in assistant.toml.

Alternatives and tradeoffs

You could paste raw screenshots directly into the Zed assistant panel. That works for one finding. Past three, the panel fills with disconnected images, your descriptions lose track of which image they belong to, and there's no record outside the editor for a teammate to read later.

You could type each finding into the assistant by hand. That's slow and the act of typing makes you summarize, you lose the texture of how the bug actually behaves. The patches land approximate.

You could file Linear or Jira tickets and link Zed to a ticket URL. That works if your team's primary feedback artifact is already a ticket. CobaltCapture is for the messier reality where the feedback exists but no one's filed a ticket yet, and the next step is "ask the assistant to fix it." For the canonical pattern behind it, see screen capture to markdown.