review-video-with-pingfusi

Review a video with pingfusi

A machine can render a video; it cannot tell you whether the result lands. This skill puts the video in front of a real human reviewer who scrubs it, pins comments to exact timestamps, draws on frames, answers your questions, and returns a verdict — then you fix the source, re-render, and refile until it passes.

Two shapes, one tool. Matching your own render against a brief you wrote is the iteration loop it was built for. But the brief is optional: point it at any video and ask your own questions when there is no prompt behind it to match.

Non-negotiables

• Publish before review. The reviewer is remote: video_url must be a public, long-lived MP4 whose host answers Range requests with 206 + Content-Range (the service probes it at file time and refuses the round otherwise). A new render is a new URL — never mutate the bytes behind a URL a round already cites.
• The brief must be honest, when there is one. current_brief is what the video must match NOW. Superseded prompts go into prompt_history marked replaced — never silently dropped; the reviewer resolves conflicts by state, not guesswork. requirements are concrete, checkable claims, each naming the prompt_ids it came from. All three are OPTIONAL: a video you did not generate from prompts you control — a competitor's ad, a tutorial, a clip someone sent you — has no brief, and you say what to judge with video_intro and steps instead.
• Everything except title is private until claim. steps, verdict_options, video_headline, video_intro, the brief, the history and the requirements all travel in a payload delivered to exactly one reviewer when they take the job, so a question may quote the brief it belongs to. title is the opposite: it is the PUBLIC headline on the row every reviewer sees while browsing, before anyone claims. Put nothing in it you would not publish.
• Ask what you actually want to know. Omit steps and verdict_options and you get the generic prompt-match questionnaire with Matches the prompt / Needs another pass, which is right for a render-against-brief loop and wrong for almost everything else. A round asking "would you keep watching past five seconds?" tells you something the generic pair cannot.
• Act on feedback in the SOURCE. A timestamped comment means a fix in the composition code, the prompt, or the asset that produced that moment — never a hand-patched frame or a trimmed clip to dodge the note.
• Never approve your own render, and never infer approval from prose. Done is a fresh core.review.verify(stateFile) returning ok === true on the declared verdict.