CobaltCapture
Draft. This page is a starting point — it needs your edit and sign-off before going public.

Feedback for AI coding agents

Capture what's wrong, talk through the fix, and hand your agent a markdown brief it can paste straight into Claude Code, Cursor, or Codex.

The problem

Coding agents are good at acting on specific, fixable feedback. They are bad at acting on "this page feels off." Most product feedback today still arrives as a Slack message with a screenshot and a paragraph that mixes the bug, the cause, and the wish-list into one wall of text. A person can parse that. An agent gets lost halfway through.

The fix is not to type harder. The fix is to capture the screen the moment something is wrong, narrate the problem out loud, and ship the agent a structured artifact it can read in one pass.

How CobaltCapture works for this

  1. Pop out the capture window. Cobalt floats on top of whatever product you're reviewing — no app switching, no losing your spot.
  2. One click per screenshot. Capture exactly the part of the screen that's wrong. Crop it down if you want.
  3. Dictate the commentary. The browser-native dictation turns your spoken explanation into editable text right next to each screenshot.
  4. Finish. You get a public URL and a markdown export. Hand either to your agent.

Why markdown beats screenshots-in-Slack

  • The agent can actually read it. Markdown with embedded image URLs is the format every code-aware LLM is trained to understand.
  • The screenshots are URLs, not attachments. The agent fetches them as part of its context — no copy-pasting one image at a time.
  • The commentary is text, not voice. "Move this button up" in dictated text survives the model's context window. A 90-second video does not.
  • Structure is implicit. Each screenshot has its own comment block. The agent doesn't need to guess which sentence applies to which image.

What to do with the markdown

The export is plain text. Paste it directly into:

  • Claude Code, Codex CLI, or Cursor's agent — drop the markdown into the chat or task description; the agent fetches the screenshots and proposes a patch.
  • A Linear or GitHub issue — the markdown renders inline, screenshots embed, the engineer or the agent picks it up from the issue.
  • An LLM chat — for general-purpose models, just paste and ask for the change.

Get started

Start a capture

Sign in with Google. Free.