--- name: generated-video-quality description: Check AI-generated footage for the failure modes generation causes rather than deterministic editing -- identity drift, generation artifacts, motion that breaks physics, temporal discontinuity at stitch points, lip-sync drift, and prompt adherence. Use alongside video-quality for any candidate assembled from fal-ai or google-ai clips; never in place of it. --- # Quality checks specific to generated footage `video-quality` covers what applies to every video this product makes: decode/duration/dimensions, black frames and freezes, legible text, a coherent edit, and a passing `quality-report.json`. Run it first and in full -- this skill does not repeat it. What this skill adds is the failure modes a camera cannot produce, because they only exist in generated footage: a face that drifts across shots, a sixth finger, a background that reflows between frames, motion that snaps rather than continues across a cut. A technical pass and a coherent-edit pass both miss these, because nothing about them fails to decode, and nothing about them is a black frame. Add these checks to the **same** `quality-report.json` `checks` map `video-quality` writes, alongside `technical`/`visual`/`audio`/`content`/ `captions`/`promise` -- one report, one file, not a second document to reconcile: ```json "checks": { "character_consistency": {"status": "pass", "evidence": ["officer.png vs shots 3,7,12: consistent"]}, "generation_artifacts": {"status": "pass", "evidence": ["no extra limbs, warped hands, or gibberish text observed"]}, "temporal_coherence": {"status": "pass", "evidence": ["motion continues across every stitch point checked"]}, "motion_plausibility": {"status": "not_applicable", "evidence": ["no complex physics in this cut"]}, "lip_sync": {"status": "pass", "evidence": ["dialogue in shots 2,5 reads in sync"]}, "clip_color_consistency": {"status": "pass", "evidence": ["black level/white balance consistent shot to shot"]}, "prompt_adherence": {"status": "pass", "evidence": ["each shot matches its shot-list framing and camera move"]}, "narration_alignment": {"status": "pass", "evidence": ["visuals match narration-manifest.md's measured durations"]} } ``` Same discipline as the base contract: every check here is `pass` or genuinely `not_applicable` with evidence, never `not_applicable` to hide a check you didn't run. A candidate with no recurring character still runs `character_consistency` as `not_applicable` -- it does not get skipped. ## Character/subject consistency The characteristic defect of this whole route, and the one most likely to be missed, because it is invisible unless checked against the actual reference rather than from memory. For every recurring character, presenter, or product: - Open its reference image from `characters/` (see `video-cinematography`) beside every shot it appears in, side by side -- not sequentially, not from what you remember the reference looking like. - Compare face, build, and the exact outfit/markings the character spec named as disambiguating. A close match on everything except the one attribute the spec called out to distinguish this character is still a fail -- that attribute existed because it is what a viewer would notice. - Where a shot conditioned on the reference still drifted, that is a generation failure to report, not a QA nuance to soften. ## Generation artifacts Look for the failure signatures specific to generative models, which a general "does it look right" pass tends to miss because each artifact is small in isolation: - warped or extra fingers, merged limbs, an incorrect number of limbs; - a face that morphs or drifts within a single continuous shot, not just across cuts; - background elements that shift, duplicate, or reflow between frames; - rendered text that is gibberish, or a logo that is almost-but-not-quite the real one; - an object that changes material, color, or shape mid-shot with no in-story reason. ## Temporal coherence at stitch points A cut between two independently-generated clips can be technically clean (matching codec, resolution, frame rate -- see `video-editing`) and still fail here, because the cut point is exactly where two separate generations meet with no shared motion between them: - At each cut, does the subject's position, pose, and motion direction continue plausibly from the last frame of the outgoing clip to the first frame of the incoming one, or does the subject visibly jump or teleport? - Does lighting and color grade hold across the cut, beyond the technical color-consistency check below -- a technically-matched clip can still read as a different scene. - Where `video-cinematography`'s overlapping-prompt technique was used to smooth a seam, check whether it actually worked at that specific cut. ## Motion and physics plausibility `video-cinematography` already warns that complex physics (explosions, crowds, collisions) is where generation models are least reliable. Check specifically for it rather than assuming a smooth render means correct motion: - Does contact between objects look physically grounded -- a hand actually gripping, a foot actually landing -- or does something pass through something else? - Is there garbled or physically incoherent motion during any fast or complex action, especially group scenes? - Mark `not_applicable` with a one-line reason for a cut with no complex physics, rather than leaving the check unexplained. ## Lip-sync drift For any shot with generated dialogue (native model audio or lip-synced narration): - Confirm mouth shapes track the actual words at a few points across the shot, not just at the start. - Flag drift that worsens over the shot's duration -- some models sync well at the open and drift by the close, which a single-frame check would miss entirely. - Mark `not_applicable` for shots with no on-screen speaker. ## Per-clip color and exposure consistency `video-editing` describes the fix (a light, consistent grade pass across the assembly) but nothing verifies whether it actually worked. Confirm it did: - Sample a frame from each clip in a scene and compare black level, white balance, and saturation against its neighbors. - A mismatch here often means two clips came from different models or providers -- cross-check against `production.json`'s per-shot model record, and note it as a scope decision (per `video-cinematography`'s one-model-per-arc rule) rather than only a color-grade problem. ## Prompt adherence Compare each shot against what `SHOTLIST.md` actually asked for -- a plausible-looking shot that quietly substituted a different framing or camera move is easy to wave through if you only check that the shot looks good in isolation: - Does the framing (wide/medium/close-up) match what was specified? - Does the camera movement match the named primitive, or did the model default to something else -- a static shot when a push-in was asked for reads as a miss even though nothing about it looks wrong? - Where a shot substituted something else, decide whether the substitution still serves the beat (accept and note it) or needs a re-prompt (per `video-cinematography`'s iteration strategy) -- do not silently accept a drift from the plan without recording that a decision was made. ## Narration and visual alignment For a narrated piece, confirm the assembled visuals actually match `narration-manifest.md`'s **measured** durations (see `video-editing`'s narration-first assembly), not the shot list's planned ones: - Does each visual segment's actual length match its narration segment's measured length, not an estimate that was superseded? - Is there a beat where narration and visual reference different content -- narration describing an action the corresponding shot does not show? ## Where this fits - Run `video-quality` first, in full -- this skill adds to that pass, it does not replace any part of it. - Read `video-cinematography` for what each character/consistency/motion rule was trying to achieve, so a finding here can be traced to the production decision that caused it. - Read `video-editing` for the stitching and grading mechanics whose results this skill verifies. - A failing check here is fixed the same way any QA failure is fixed: the smallest responsible layer re-generates or re-cuts, not the whole piece.