You captured a screen, dropped a pin on the broken button, and typed what's wrong. Then a developer fed your review to an AI coding agent, and the fix came back touching the wrong element. The gap is usually vocabulary: you and the agent mean different things by the same words, or you handed it input it could not read.
Four terms cover most of that gap. Each one changes how you write the review before you send it. For the complete reference, the vibecoding glossary runs deeper, but these are the four that touch your feedback directly.
Prompt: the instruction the agent acts on
A prompt is the text instruction you or a developer give an AI coding agent to make it do something. "Fix the overlapping labels on the pricing table" is a prompt.
This matters because your review often becomes the prompt, or most of it. When an agent reads a Cobalt review over its MCP connection, your pinned screenshot and your comment are the instruction. If your comment says "this looks off," the prompt is vague and the fix will be too. If it says "the Save button sits half under the footer on the 25-seat plan," the agent has something specific to act on. Write the comment as if it were the instruction, because it is. The feedback specific enough to act on post walks through what to include.
Turn: one round of ask and response
A turn is a single exchange with the agent: you ask, it responds, and whatever it changed is the result of that turn. A fix that lands "in one turn" means the agent got everything it needed the first time and did not have to come back asking what you meant.
Turns are the unit that wastes time. Every round trip where the agent asks "which button?" or "what should it look like instead?" is a turn spent on clarification instead of a fix. A review with a cropped still, a numbered pin, and a plain comment is built to be answered in one turn. If your feedback forces three turns of back-and-forth before the first code change, the feedback was the bottleneck. See a bug report Claude Code fixes in one turn for what that looks like in practice.
Context: what the agent can actually see
Context is the information the agent has in front of it while it works: your comment, the screenshot, the relevant code, and whatever else fits in its working memory for that session. The agent cannot act on anything outside its context, no matter how obvious it is to you.
This is why a bare link to a live page often fails. If the agent cannot open the page, or the page needs a login, the thing you are pointing at never enters its context. A captured still carries the visual into the context directly. So does markdown with the image inline, which is why Cobalt serves the markdown version of a review automatically when an assistant opens the link. The rule of thumb: if it is not in the screenshot or the text, assume the agent does not know it.
Agent-readable artifact: feedback in a format an agent can parse
An agent-readable artifact is a piece of feedback structured so an AI agent can read it start to finish without a human translating first. Clean markdown with headings, the screenshot inline, and one point per item is agent-readable. A screen recording is not, because the agent cannot watch it.
This is the distinction that decides whether your review gets used or re-typed. A review published through Cobalt Capture is readable by people and by agents at once: the /r/<slug> link opens as a page for a human, and /r/<slug>/markdown gives an agent the same content as plain text with images it can parse. You do not pick one audience. The what agent-readable feedback means guide covers the format in full.
How the four terms work together in one review
Put them in sequence. You capture a screen, crop it to the broken region, pin the exact spot, and write a comment that reads like an instruction. That makes the review an agent-readable artifact. When a developer hands it to their agent, your comment becomes the prompt, the screenshot enters the agent's context, and because nothing is missing, the fix lands in one turn.
Skip any step and the chain breaks. No pin, and the agent guesses which element you meant. A live link instead of a still, and the visual never reaches the context. A vague comment, and the prompt is weak no matter how sharp the screenshot is.
None of this requires you to use an AI coding agent yourself. Most reviewers do not. You capture the screen in a browser tab, talk or type through it, and publish a link. There is no install, no extension, and no password, and you can try it free before anything is paid. If a review sits untouched for 7 days it goes read-only for you while the shared link keeps working, and Pro at $5 a month keeps everything editable with no cutoff.
Start with one captured still, one pin, and one comment written like the instruction it will become. Capture a screen and write your first review, then read it back and ask whether an agent could fix it without a single follow-up question.