Someone on your team says a review is "agent-readable" and you nod, but the term is doing a lot of quiet work. It has a specific meaning, and getting it wrong wastes a round trip with your coding tool. Here are the terms defined, each with a one-line definition and a note on when the distinction actually costs you time.
Agent-readable feedback
Feedback packaged so an AI coding agent can parse it without a human retyping it first. In practice that means plain text: a heading, a short description of what is wrong, where it is on the screen, and what you want instead. No video the agent cannot watch, no screenshot buried in a chat thread with no words attached.
This matters the moment you hand work to a tool like Cursor or Claude Code. If your feedback is a screen recording and a spoken "the button looks off," the agent gets nothing to act on. It needs the words and the location as text. For the broader vocabulary around this, the vibecoding glossary collects the related terms in one place.
The parts of the term
Cropped still
A single frame captured from your screen and trimmed to just the part you mean. It is not video and it is not an annotated live page. You drag a rectangle around the broken component and drop everything else. The crop is the only image edit you make, which keeps the reference unambiguous.
A cropped still beats a full-page screenshot for the same reason a short sentence beats a paragraph: the receiver spends zero effort figuring out what you meant. There is more on why a cropped still works better than marking up the live page if you want the reasoning.
Markdown export
The plain-text version of your review, structured with headings and lists, that an agent reads directly. Where a person opens the review as a link or a PDF, an agent opens the markdown. Same content, format matched to the reader. If markdown itself is fuzzy for you, the terms behind markdown screenshots spell it out.
Pin
A numbered marker placed on a captured still that points at an exact spot. Instead of writing "the icon near the top right," you drop pin 2 on the icon and your comment refers to pin 2. The agent maps the number to the coordinate. This removes the guesswork that pinning the exact spot is designed to solve.
A short worked example
Say a form's submit button stays disabled after you fill every field. You capture the form, crop to the button and the last input, and drop a pin on the button. Then you type a comment. The person-facing version reads as a normal review item. The markdown version an agent pulls looks close to this:
## Submit button stays disabled
Screenshot: signup form, pin 1 on Create account button.
Expected: button enables once email and password are valid.
Actual: button stays greyed out after both fields pass validation.
That is agent-readable. It names the screen, points at the element, states expected versus actual, and carries the image reference. An agent can turn it into a change without a follow-up question. A human reading the same review gets the identical facts in a nicer layout.
What agent-readable is not
It is not a bug tracker with statuses and assignees, and it is not a screen recording. A recording is time you make someone spend watching; the agent cannot watch it at all. If you have been sending video and wondering why the round trips drag, the reasons a recording fails as agent input cover it directly.
The practical test is simple. Can the receiver, human or agent, act on what you sent without asking you a question? If yes, it is agent-readable. If they have to reply "which button?" or "what did you expect?", it is not, and you are about to lose a day to clarification.
Where the output comes from
Cobalt Capture runs in a browser tab with no install, no extension, and no signup. You capture a screen, crop, pin, and comment by typing or dictating, then publish. One review gives you a public link, a PDF or Word document, and the markdown an agent reads, all from the same source. When you want to try the flow on a real screen, start a new review and publish it in a couple of minutes.