You captured a broken header on staging, wrote two sentences about it, and pasted the result into Cursor instead of Slack. The receiver is software. That single fact changes what your feedback needs to contain, what format it arrives in, and what counts as enough detail. The vocabulary below is the small set of terms that come up once a coding agent is on the other end of your review.
The artifact and its formats
Feedback artifact. The single output of a review session: a set of screenshots, comments, and pins, addressable by URL and exportable to several formats. When a person reads it, they open the public link or the PDF. When an agent reads it, it ingests the markdown version of the same artifact. One source, different renderings.
Agent-readable feedback. Feedback in a format the model can actually parse: text, with image references as URLs, structured by headings and lists. A Loom video is not agent-readable. A PNG with arrows drawn on it is partially readable at best. A markdown file with inline image links is. For a longer take, see what counts as agent-readable feedback.
Markdown export. The plain-text version of a review served at /r/<slug>/markdown. Each item becomes a heading or list entry, each screenshot becomes an inline image link, each pin becomes a numbered reference inside the comment text. This is the URL you hand to an agent. The markdown screenshots format guide walks through exactly what the output looks like.
The pieces inside the artifact
Item. One unit of feedback. Usually a screenshot with a comment attached, sometimes a floating comment with no image. Each item should hold one observation. "Header overlaps the search input at 1280px and the CTA button has the wrong hover color" is two items, not one. An agent will try to act on each item independently, and a mixed item tends to get a mixed fix.
Pin. A numbered marker placed on a screenshot pointing at a specific spot. Pins exist because prose is bad at coordinates. "The 24px gap above the second card" is fine; "the third icon from the left" is ambiguous. A pin labelled 1 lets you write "1: this icon should be the filled variant" and the agent reads the number alongside the image, with no parsing of left-right counting required.
Crop. The rectangle you drag on a captured still to narrow what the receiver sees. Cropping is the only image edit in Cobalt Capture; there is no drawing of arrows or labels on the screenshot itself. For agents that read screenshots as input, a tight crop reduces the irrelevant pixels the model has to reason about. For a person, it just makes the issue obvious.
Comment. The text attached to an item, typed or dictated. This is the part the agent treats as instruction. Vague comments ("this feels off") produce vague diffs. Concrete comments ("replace the 32px top margin with 16px on viewports below 768px") produce something you can review.
Terms for the workflow around the artifact
Public review link. The short URL of the form /r/<slug> created when you publish. The same link works for a person opening it in a browser and for an agent fetching the markdown variant. One review, two consumers, no second export step. See one review serving the client and the developer for the pattern in practice.
Screenshot-to-prompt. The act of turning a captured screen into something you paste into an agent's chat. The naive version is dropping a PNG and saying "fix this." The useful version is a markdown block with image, pin numbers, and instruction prose. The screenshot-to-prompt workflow covers the mechanics.
Handoff. The moment a human review becomes an agent's input. In a mixed team you might do the capture and triage yourself, then paste the markdown URL into Claude Code or Cursor and let it propose patches. The human-to-agent handoff page describes where the boundary sits.
Resolved. A state on each comment, set by the review owner once a fix lands. Resolved is the closest Cobalt Capture gets to workflow; there are no assignees, no statuses, no sprint columns. The agent does not read or write this state. It exists so you, the human, can track which items have been actioned without leaving the review.
Why the vocabulary matters when the reader is software
A person tolerates ambiguity. They squint at a screenshot, guess what you meant by "the button," and ask if they are wrong. A model does not ask. It picks an interpretation and writes code against it. That is the whole reason feedback aimed at an agent deserves its own terms: item, pin, crop, and comment are the levers you actually have for forcing precision.
The shortest path to trying it is to capture a real screen, add a pin and a sentence, publish, and append /markdown to the URL before pasting it into your agent of choice. Start a review and look at what comes out. The structure of the markdown will tell you more about what your agent needs than any glossary can.