You build a screen with an AI coding agent, look at the result, spot something wrong, and tell the agent how to fix it. That round trip is the feedback loop, and in a vibecoding context the word carries a precise meaning. If you review work that an agent built, these are the terms you will trip over. Each one below gets a one-line definition and a note on when it actually matters.
Feedback loop
Feedback loop: the cycle of looking at what the agent built, describing what is wrong, and feeding that description back so the next version is closer.
In older workflows the loop was you, a code diff, and a text editor. In vibecoding the loop is you, a rendered screen, and a prompt. The screen is where the problems show up, so the quality of what you feed back decides whether the next turn fixes the bug or wastes it. A vague "the header looks off" starts a slow loop. A cropped screenshot with a pin on the exact element and one sentence of what you want starts a fast one. The whole point of a feedback loop for vibecoding is to keep each turn short and specific so you ship instead of circling.
Still
Still: a single captured frame of a screen, a static image, not video.
The still is the raw material of visual feedback. You click Capture screen in the browser, the browser asks which window to share, and the current frame is grabbed as an image. On a phone, where browsers cannot capture the screen, you take a normal screenshot and add that instead. The reason a still beats a recording here is that an agent cannot scrub a video. It reads a frozen frame, and everything you mark on that frame stays put. There is no scrubbing to the right moment and no guessing which second you meant.
Pin
Pin: a numbered marker you drop on a specific spot of the still to point at exactly what you mean.
Pins matter most when the screen is busy and "the button" could be any of six buttons. A pin labeled 1 on the misaligned CTA and 2 on the overlapping label removes the ambiguity. The numbers also set reading order, so the person or agent reads your points in the sequence you intended. Pins are baked into the exported image, so whoever opens the review sees the marks without needing your tool. If you want the full vocabulary for aiming at parts of a screen, the words to use when pointing at a button or region cover it in depth.
Item
Item: one unit of feedback, usually a screenshot plus a comment, though it can be a free-floating note with no image.
A review is a list of items. Each item pairs a marked still with what you typed or dictated about it. This is the structure that makes feedback actable: instead of one long paragraph that mixes five problems, you get five discrete items, each pointing at one thing. A developer or agent can work the list top to bottom and mark each one done. Keeping items small and single-topic is the difference between feedback someone fixes and feedback someone argues with.
Review link
Review link: a short shareable URL of the form /r/<slug> that anyone can open to read your review, no login required.
When you publish, the review gets this link plus a clean reader page at /s/<slug> for people who only need to read. The receiver opens it in a browser. No install, no extension, no account. People with the link can comment on any individual item, and you as the owner can mark each comment resolved. That is why a link beats a pile of attachments: the conversation lives next to the exact screen it is about. What your client sees on a review link is the plain reader view, nothing to figure out.
Markdown URL
Markdown URL: the same review served as plain-text markdown at /r/<slug>/markdown, the format a coding agent reads.
This is the quiet workhorse of the agent loop. A human opens the review link and sees screenshots and comments. An AI assistant that opens the same link is served the markdown automatically, screenshots included, so it reads your pins and notes as structured text it can act on. You do not maintain two documents. One review feeds both the person and the agent. If you want the mechanics, the piece on review link or markdown URL for your agent walks through which one to hand over.
Where these terms sit next to each other
Run them in order and the loop reads cleanly: you grab a still, drop a pin, write the item, publish the review link, and hand the agent the markdown URL. Each term is one step of the same round trip. Get the vocabulary straight and the loop stops being a fuzzy idea and becomes a procedure you can repeat.
These five are the ones a reviewer hits first, but they are not the whole set. The full vibecoding glossary defines the rest, from handoff to agent-readable feedback, in the same plain terms.
Cobalt Capture is free to try with unlimited reviews and no password; after 7 days a review goes read-only for you while the shared link keeps working, and Pro at $5 a month keeps everything editable. Start a review and capture your first still to see the loop end to end.