Reference Video Analysis
Original, project-agnostic procedure. Agent-host effectiveness remains experimental.
When to use
Inspect authorized reference footage with timecoded visual and audio evidence, then turn its structure into a distinct adaptation brief rather than guessing from a transcript.
When not to use
Do not impose this procedure on unrelated small tasks. It does not install tools, create a worker, authorize spending or publication, or replace the consuming repository’s rules. Private paths, account bindings and accepted policy remain outside this public package.
Inputs
Resolve the actual task, accepted source/artifact identity, authorized scope, installed capabilities, output requirements and evidence available. Read existing project configuration before asking the user to repeat it. Identify missing prerequisites rather than fill them with invented observations.
Procedure
1. Separate inspiration from source editing
Determine whether the user wants a structural reference, an edit of owned footage, or both. Confirm the available local file or approved access method and allowed use. Record the actual media duration, timebase and language. Do not download restricted footage or assume an accessible link grants reuse rights.
2. Inspect the modalities independently
Watch representative intervals and inspect transition-adjacent frames; read available captions and listen to authorized audio separately. Record timestamp, observation and evidence modality. A transcript establishes spoken text, not framing, motion or visible product behavior. Sparse frame sampling cannot prove an entire action or every cut; state the sampled coverage.
3. Describe mechanics rather than borrowed content
Map the opening promise, information sequence, pacing, visual hierarchy, shot-scale changes, typography and audio relationships. Separate measurements from interpretations of why the piece works. Note uncertainty from variable frame rate, imprecise seeking, missing audio or cropped captures. Avoid identifying real people from images.
4. Develop an original adaptation
Propose a small set of concepts that reuse abstract structural ideas but use an independently written script, approved assets and a distinct composition. Explain what changes for the target audience, duration and aspect ratio. Never lift another creator’s script, character identity, logo or personal endorsement by default.
5. Supply a production-ready brief
Hand Product Video Storyboarding a timecoded observation table, original message hierarchy, proposed beats, source restrictions and unresolved evidence. Keep reference footage separate from approved production assets. The analysis itself does not authorize paid generation, cloning a person’s voice or publishing a video.
Output
A task-sized artifact and review record with input identities, decisions, source-linked observations, actual check results and limitations. Keep implementation, supplied metadata, measured behavior and independent acceptance distinct. Use passed, failed, blocked and not-inspected states rather than a blanket success claim.
Failure handling
Only captions are available. Produce an audio/text structure analysis and mark visual pacing, framing and continuity unevaluated instead of inventing what the video shows. Preserve partial work and explain the smallest missing input or corrective step. Never broaden tool authority to conceal a blocker.
Example
A supplied 24-second demonstration uses a rapid problem statement, three feature beats and a quiet closing card. The analyst verifies those visual beats in the file and proposes a 15-second original explainer using synthetic shapes, not a copy of the script or footage. This is a synthetic example, not a production or agent-host result.
Companion and evaluation
See collection contracts and worked fixtures. The helpers check only their documented input contracts. references/scenarios.json contains not-run host evaluation inputs, not recorded executions.