One frame from a small Twitch stream, treated as the preview thumbnail a viewer scrolls past in the directory: what is this, would anyone click, and what would fix it. The single-frame sibling of stream-audit-consultant with the costume off — no clip to read motion from, no consultant persona to hide behind, and a prompt that says out loud there is no sound, which is the trap.
The single-frame sibling of stream-audit-consultant, and worth running back to back with it: same channel, same perception task, one still instead of a clip. Everything a model could infer from motion or sound is gone, so what is left is what it can actually see — and a frame off a live stream is the thumbnail, which makes "would anyone click this" the honest question rather than a framing device.
Ground truth is whatever frame you grabbed, so score it with the picture open and a finger on the screen.
The read, and the Evidence line under it. "What this is" should be about this stream — the game or the activity, what is on the second monitor, what the streamer appears to be halfway through. Read the Evidence line first: every item on it should be pointable-at. That line is the whole design of the section, because it turns a confident guess into a checkable one. Rank: right read on evidence that holds > wrong read on evidence that holds (it looked at the picture and reasoned badly, which is the smaller failure by a long way) > right read on evidence that isn't in the frame (it recognized the game from the shape of the HUD and then decorated) > possibilities in a trenchcoat, which failed the instruction before it was scored.
The audio trap. There is no audio here, the prompt says so explicitly, and models will review the mic anyway. This is the still set's version of the clip set's audio-honesty axis and it carries the same weight: a line about the streamer's energy or their music balance is fabricated from the concept "Twitch stream" rather than from the picture in front of it, and it outranks any amount of good advice as a fault. Watch for the hedged version too — "the pacing feels relaxed", "they come across as engaged" — which is the same failure wearing a qualifier.
Fabrication generally. One invented object sinks the answer no matter how well the rest reads: the second person who isn't there, the chat message it quotes that nobody sent, the sponsor logo read off a four-pixel banner. Overlay text is the common one — a model will quote a follower alert or a recent-sub name at a resolution where it cannot possibly have resolved the letters.
You are an experienced social media strategist who has spent years growing live channels. In front of you is a single still frame from a small Twitch stream.
Treat that frame as the preview thumbnail — the one a viewer sees while scrolling the Twitch directory, with about a second and a half to decide whether to click. Everything you say has to come out of this one image.
You received one still. There is no audio and no motion. Do not comment on voice, mic quality, music, energy or pacing, or on anything else you would have to hear or watch over time to judge. If a section pulls you toward it, describe what the frame shows instead.
Write plainly. No comedy, no persona, no jargon for its own sake — the way you would talk to a client paying for the half hour. They are starting out and trying hard, so be straight with them rather than harsh, and say what is genuinely working.
Twitch's own interface is not the streamer's work: ignore the Follow / Subscribe / Gift a Sub buttons, the player chrome, the viewer count and any Twitch-served ad panel. Their own work is fair game — overlays, webcam frame, alert graphics, banners, branding — and so is the chat, if messages are readable. You may read the channel name and use it.
Answer in exactly these four sections, in this order.
What is this stream about, and what is happening in it at the moment this frame was taken? Commit to one reading; do not list possibilities. Two or three sentences, then a final line beginning "Evidence:" naming the specific things in the frame you read that from.
Layout, framing, lighting, overlays, readability, color, branding — what a first-time viewer takes in. Two or three sentences about this frame in particular, not about streams in general. Name one thing that is genuinely working.
Begin with exactly one of these three words, alone on its own line: Click, Maybe, or Scroll past. That is your honest call on whether someone browsing the directory stops on this thumbnail.
The verdict. One of three words, alone on a line. Hedging it into a sentence is a broken format instruction. Answering "Maybe" is legitimate once and a tell across a session — that one is worth tracking over several frames rather than marking in a single column. The call itself is not really the scored part, since reasonable people click on different things; the reasons are, and they have to be about this thumbnail. "Scroll past" argued from something visible beats "Click" argued from nothing.
The three fixes. This is where generic advice goes to hide, and generic is the failure mode here, not harshness. "Improve your lighting" fits any channel; "the only lamp is behind you, so the webcam is a silhouette against the window — move it to the left of the desk" is a fix. Count them: exactly three, ranked. A model that offers five is not being generous, it is declining to prioritize, which was the job.
The Twitch-chrome rule, which is a scored instruction. Listing the Follow button, the player controls or the viewer count as a problem is a read failure: it criticizes Twitch's product as though the streamer had built it, and the prompt names those elements explicitly. It is also the tell that a model is auditing the screenshot rather than the stream — mark it the way you would mark a broken output format.
Tone, which is the reason this is a separate challenge rather than a re-crop. stream-audit-consultant hands the model a persona to hide inside: a model with nothing to say can perform the jargon and sound like it has plenty. This prompt takes the costume away and asks for plain professional prose, so what is left over is whether it saw anything. A model that reaches for the bit regardless — the growth-consultant voice, the jokes, the knowing throat-clearing — is telling you it had less to say than the clip version made it look.
Then two or three sentences on why, and one sentence on what a viewer who clicks would expect to be watching a minute later.
Exactly three changes, most important first, as a numbered list. Each is one sentence that names something specific in this frame and says what to do about it. General streaming advice does not count — if a fix would fit any channel, it is not one of your three.