Shot-by-Shot Analysis

Purpose

Turn multiple reference videos into one evidence-backed Markdown report that a human can review and downstream generation skills can use. Analyze first; synthesize only after every source has been inspected.

Execution Boundary

• Use media-analysis for actual video understanding. Analyze each source in a separate request with the same canonical beat-evidence contract.
• Do not generate images, videos, prompts, workflows, or execution requests.
• Do not infer video content from metadata, filenames, transcripts, thumbnails, or general knowledge when the footage itself has not been inspected.
• Keep raw hosted responses and temporary request records internal. Deliver one Markdown report named <case>-shot-by-shot-analysis.md.

Workflow

1. Inventory every input video. Record a stable source label, path or URL, duration when known, and source basis.
2. Inspect each video independently. Do not expose earlier analyses or a draft shared style theory while analyzing later videos.
3. Segment by observable shot or meaningful continuous beat. Use start-end timecodes. Mark approximate boundaries when exact timing is uncertain.
4. Complete the beat-evidence blocks for the current video before interpreting its broader purpose. Describe visible and audible evidence before inference.
5. Verify that the blocks cover the full usable duration, preserve source order, and do not invent speech, sound, motion, products, or production intent.
6. Repeat until every video has complete beat evidence.
7. Compare all completed evidence sets. Separate recurring grammar from one-off details and unique recognizable elements.
8. Write the Style Grammar Report with concrete rules, supporting source and timecode references, recurrence, and confidence.
9. Add a compact downstream handoff section. Preserve the style logic without turning it into final prompts or execution requests.

Beat Evidence Contract

For each source, preserve the complete media-analysis Section 4 Shot-By-Shot Evidence blocks. Use references/report-template.md for the exact fields; do not convert them into a wide table or omit any field.

Apply these rules:

• Describe what is visible and audible before interpreting it.
• Keep visual evidence factual. Put inferred narrative or commercial function only in production purpose.
• Describe framing with production terms such as shot size, angle, headroom, subject placement, depth, foreground, and negative space.
• Distinguish camera movement from subject movement.
• Quote speech or visible text only when legible or audible. Otherwise write unclear, partially audible, or another precise uncertainty marker.
• Describe edit rhythm using observable cut length, cut type, pauses, speed changes, montage density, or continuity behavior.
• Mark inferred emotional state or production purpose with confidence when the evidence is ambiguous.
• Use not present when a field truly does not occur. Do not fill absence with speculation.
• Use compact tables only for short comparisons. Dense observations and quoted language stay in vertical beat blocks so the report remains readable on narrow screens and in agent context.

Synthesis Rules

• Do not average the videos early. Synthesize only from completed beat evidence.
• Treat a pattern as recurring only when supported by more than one source, or by repeated independent examples within a source. State the narrower basis.
• Cite evidence as <source-label> @ <timecode>.
• For every grammar dimension, distinguish:
  • shared rule: the reusable production behavior
  • evidence: supporting source and timecode references
  • recurrence: recurring, source-specific, or uncertain
  • confidence: high, medium, or low
• Prefer concrete language over labels such as premium, cinematic, authentic, or native. Use those labels only after describing the visible and audible evidence that supports them.
• Preserve camera grammar, timing, object logic, relationship logic, sound behavior, and emotional progression without preserving creator identity.

Imitation Boundary

Do not copy or recommend copying:

• creator identity, face, voice identity, or distinctive personal mannerisms
• exact scenes, locations, wardrobe, dialogue, jokes, or shot sequences
• brand marks, packaging identity, private details, or recognizable background information
• unique overlays, signature graphics, or other source-identifying elements

Translate useful reference evidence into general production rules. Put recognizable or source-specific elements under avoid, not must_keep.

Markdown Deliverable

Produce one report from references/report-template.md with complete source beat evidence, the ordered Style Grammar Report, and a concise handoff.

Downstream Handoff

Keep the handoff concise and derived from the report:

• Image Generation: Extract static framing, subject state, scene, lighting, palette, product placement, must_keep, can_vary, and avoid. Hand this to creative planning before image-batch-runner.
• Video Prompt Architecture: Extract camera language, movement, edit rhythm, sound, speech posture, emotional progression, product behavior, continuity, and generator risks before video-batch-runner.
• Workflow Creation: Provide the full report as source evidence. Compile its rules into positive, executable scene carriers and explicit reference jobs; do not paste analysis language or imitation warnings into model prompts.

Do not create separate handoff files unless the user explicitly requests them.

Quality Check

Before handoff, confirm:

• every input video has complete, ordered beat evidence
• synthesis begins only after the final source evidence set
• shared claims cite supporting sources and timecodes
• recurring and one-off details are separated
• uncertainty is visible rather than silently resolved
• production-purpose inference is not presented as observation
• unique or recognizable identity, brand, and private details stay out of reusable rules; the report remains human-readable and production-specific

Stop Conditions

• Stop when a source video cannot be accessed or actually inspected.
• Stop when missing sources would materially change a requested cross-video synthesis.
• If the hosted analysis command fails, report the exact error and do not replace video evidence with guesses.