--- name: video-qa description: Independently review an existing edited video against Video Studio's quality standards and report timestamped defects, coverage, and verification limits. Use for video QA, checking whether an edit is up to standard, or a final quality review. For making or fixing an edit use edit-video or iterate-video-edit; for script-only review or contact-sheet extraction alone use the relevant specialist skill. --- # Video QA Review an existing rendered edit against the shared video quality standards without modifying the video. ## Contract Checklist - Identify the exact video and iteration being reviewed; do not infer versions from filenames alone. - Read the [15 video quality standards](../edit-video/references/video-quality-standards.md) after preparing the source and before inspection. - Review the full render, including narration alignment, transitions, readable text, and speech continuity; attractive stills and export validity alone do not establish quality. - Review silently using timestamped frames, transcripts, and audio signal analysis. Play specific media only when the user directly requests its playback. - Keep findings tied to observable evidence, timestamps, and numbered standards. Distinguish defects from optional taste preferences. - Report technical quality and fulfilment of the explicit creative brief separately. A clean, readable, correctly branded export may still need creative fixes. - For reference-credited scenes, compare the recreation and output sequences side by side. A shared motif or broad category does not establish reference use; unrecognizable or materially weaker transfer is a creative failure even when technical checks pass. - Do not edit media, append Studio feedback, change status, or publish as part of an independent review unless separately requested. - Keep durable review evidence separate from disposable extraction scratch; clean only scratch created by this review under the [storage lifecycle](../edit-video/references/storage-lifecycle.md). ## Workflow 1. Prepare the exact review source. - Follow workspace instructions. Run video-studio-setup only if the machine needs configuration. For a Studio video, use its accessible export listing to resolve the requested iteration and feedback; download through authenticated read_video_chunk rather than launching a player or autoplay page. Keep media bytes out of conversation output. - For a supplied local render, record its path and available version identity. Obtain the approved script or an aligned transcript, intended viewing format, relevant source excerpts, and supplied style references where available. Use connected sources before asking for missing context. An unavailable brief does not prevent checking observable defects; record which comparisons remain unverified. - Read the independent-QA section of the [storage lifecycle](../edit-video/references/storage-lifecycle.md). Record scratch outputs as they are generated, check space before dense extraction, and reuse a verified local render when available. - Preserve originals. Use an isolated review workspace and record the source identity, duration, dimensions, and review scope. A requested excerpt review must not become a full-film pass claim. - Anti-patterns: reviewing the wrong iteration, assuming the newest export is the requested version, opening autoplay media, blocking observable checks because some context is missing. 2. Read the standards and plan coverage. - Read the [shared quality standards](../edit-video/references/video-quality-standards.md) now, after setup and source preparation and before reviewing frames. Apply all 15 in context with the approved brief, reference treatments, and explicit user instructions. - Build a coverage ledger for the full duration, scene boundaries, narration beats, and relevant prior feedback. Track each standard as pass, fail, not applicable with a reason, or unverified with a specific limitation. - Identify the requested creative outcome and use the [creative-brief comparison](../edit-video/references/reference-review.md#judge-the-full-sequence-against-the-creative-brief) to plan sequence-level checks. Do not replace the user's brief with a personal preference for constant movement, a specific medium or an arbitrary number of references. - Treat stability during reading, deliberate negative space, and a reference-supported visual style as contextual choices. Do not invent a universal cut interval, animation quota, or mandatory prop palette. - Anti-patterns: judging from memory of the checklist, treating all motion as useful, converting one scene's feedback into a universal visual rule. 3. Inspect the rendered video silently. - For reference-informed graphics, use the shared [reference comparison guide](../edit-video/references/reference-review.md#compare-rendered-behavior-then-improve-it) to compare the selected recreation and rendered phases side by side. New content and theme are legitimate adaptations, but the defining composition and sequence must remain recognizable when the output claims that reference. Judge intentional original scenes on their own purpose without inventing an attribution. - When `referenceUsages` is present, use its output ranges and linked reference/recreation IDs to check the claimed adaptations. Flag ranges outside the render or claims not supported by the visible result; an empty list is not a quality failure. - Audit [actual reference transfer](../edit-video/references/reference-review.md#verify-what-actually-transferred) independently of the editor's claims. Classify each claimed mechanism as substantial, partial, absent or unverified from timestamped recreation/output sequences; distinguish studied clips, generic shared features and implemented distinctive sequences. Partial or absent transfer is not an acceptable `referenceUsages` claim. Inspect adjacent scenes for repeated heading/caption/panel structures and reveal rhythms even when their subject matter differs. - Extract timestamped contact sheets across the entire duration at roughly 0.25–0.5-second intervals, batching as needed. Inspect larger frames for small text and details; extract denser sequences around reveals, transitions, and suspected glitches. Include the first and last frames. Compare against actual reference frames where style fidelity is in scope. - Align the script/transcript with frame sequences to check when concepts appear, how long they remain readable, and whether compositions evolve coherently. Check repeated imagery, relevance, interface behaviour, chronology, evidence, clipping, spacing, and contrast against the numbered standards. - Inspect transcripts for repeated takes and false starts, and waveforms/silence/loudness/clipping measures for candidate audio defects. Separate confirmed transcript or signal findings from suspected clicks, speech glitches, and listening-dependent judgments. Never claim to have heard audio from silent analysis; mark remaining uncertainty explicitly. - Anti-patterns: reviewing only hero frames, hiding unreadable text inside a small sheet, missing transition states, declaring audio clean solely because peak levels pass. 4. Report findings and the review outcome. - For each finding record the exact iteration, time range, standard number, observable defect, supporting frame/transcript/signal reference, and concrete correction. Group duplicates when one fix applies across multiple occurrences, retaining affected timestamps. - Use “Needs fixes” for evidenced quality failures, “Incomplete review” when required coverage or material checks remain unverified without an evidenced failure, and “Pass within checked scope” only when applicable checks pass. A report with failures and unverified checks must disclose both. Do not claim publishing approval. - State separate technical and creative outcomes. If the explicit creative brief is unmet, including a claimed reference-led scene that is unrecognizable or visibly weaker than its recreation, retain “Needs fixes” despite technical passes; cite the missing composition or phase sequence rather than only calling it boring. Reassess an earlier generous pass when new comparison evidence contradicts it, retaining the prior report and explaining the correction. Editorial judgment does not establish measured audience engagement. - Prioritise misleading evidence, unintelligible or repeated speech, unreadable essential content, and broken animation/layout as blocking defects. Separate other quality failures from optional refinements. For a bounded revision, flag unrelated defects without authorizing their repair. - Save the report and coverage ledger with the review workspace. If asked to verify a later correction, recheck affected ranges, adjacent transitions, and relevant unchanged regions against the exact new render, updating the remaining findings rather than assuming fixes worked. - Anti-patterns: vague aesthetic verdicts, averaging serious failures into a passing score, unsolicited editing or Studio writes, treating QA findings as permission to redesign. 5. Retain evidence and close review scratch. - Apply the [independent QA cleanup rules](../edit-video/references/storage-lifecycle.md) only to files created by this review. Verify the saved report, coverage ledger, contact sheets, cited frames, and transcript/signal evidence first. Retain the verified source and extraction settings; remove only uncited re-extractable scratch, record reclaimed space, and check evidence links afterward. - Anti-patterns: deleting cited frames, treating repair images as extraction scratch, removing production inputs or another review's files, sweeping shared caches during read-only QA. ## Anti-Patterns - Treating a valid export as proof of visual quality. - Treating a technically clean scene with a generic split, document or color flip as a faithful recreation of a distinctive reference sequence. - Adding decorative motion to satisfy the checklist without helping the explanation. - Calling an excerpt or a sparse sample a complete review. - Claiming listening-based conclusions from silent inspection. - Changing the video or its review status during a read-only QA request. ## Output Identify the video and exact iteration, give the review outcome, and list prioritised timestamped findings with standard numbers and actionable corrections. Link the saved report and silent evidence when available. Summarise coverage, unverified checks, and optional refinements separately. Behavioral cases are in [tests/scenarios.md](tests/scenarios.md).