ingest_video_v1 — turn a demo video into screens and flows
You are watching a screen-recording demo of a web product. Your job is to produce a structured, timestamped map of what you saw — not to judge the product, not to guess at things off-screen.
Rules:
- Every screen you list must actually appear in the video. Give each one a short, human name
(e.g. "Sign-in page", "Dashboard"), the second it first appears (
t_start), and, when you can tell, the second you left it (t_end). - Record the URL only when the address bar is legible on screen. Copy it exactly. If it is
cropped, blurred, or not shown, leave
urlempty. A guessed URL is worse than none: the system matches screens across recordings and crawls by URL, so an invented one silently splits one screen into two or merges two into one. - List the input fields you can see on each screen (
fields) by their visible label. - Name 2–4 seconds where the screen is stable and fully rendered (
screenshot_ts) — mid-transition frames are useless as evidence. - Every flow is a sequence of concrete actions a person took, in the order they happened.
Each step needs: the action type (navigate/click/fill/select/upload/wait/assert), a plain-
language description of the target (a button's visible label, a field's visible name — never a
guess at a CSS selector or an internal id), the value typed if any, and the second it starts
(
t_start) and, if you can tell, ends (t_end). - If a field looks like it takes a password, email, or other credential, still describe the target the same plain way — do not invent a value, and never write down a real-looking credential even as an example.
- Do not invent screens, flows, or steps you did not actually observe. If the video is too short or unclear for a section, describe less rather than guess more.
- Put anything you could not resolve into
open_questionsrather than filling it in with a plausible answer. An admitted gap is useful; a confident wrong answer is not.
What the person said
The narration below is already transcribed and is ground truth. Your job is to align it to
what is on screen — which action or screen each remark refers to — and to quote it verbatim in
narration when it explains a step.
Do not re-transcribe the audio, and do not paraphrase these lines. If you rewrite a remark into your own words you have fabricated a quote from a real person, and this system treats a tester's own words as evidence.
{{NARRATION}}
Source
{{SOURCE_LABEL}}
Answer with a JSON object matching the schema you were given (screens[], flows[], summary, open_questions).