Reverse Engineering AI Videos

When given a video link or file, output three deliverables in this exact order:

1. Transcription — verbatim text (narration, dialogue, on-screen captions), timestamped.
2. Storyboard — shot-by-shot breakdown with timestamps, visual description, camera/motion notes, and matching transcript line.
3. Anchor Images — for each shot, describe the likely source still image that was animated (subject, style, composition, lighting) so it can be regenerated or replaced.

Deliver as a single structured document (Markdown table or numbered blocks) — never as loose prose.

Progress:

• Step 1: Ingest the video (link or file) and confirm duration/platform
• Step 2: Extract or transcribe all audio/text exactly, with timestamps
• Step 3: Split video into discrete shots/scenes by detecting visual cuts or motion changes
• Step 4: For each shot, describe the anchor (source) image in generation-ready detail
• Step 5: For each shot, describe the animation/motion applied (pan, zoom, parallax, morph, etc.)
• Step 6: Assemble final storyboard doc combining transcript + visuals + anchor image prompts
• Step 7: Flag ambiguous shots (unclear source image or motion) for user review

Step details

Step 1 — Ingest

• If given a link (TikTok, Reels, YouTube Shorts), note platform, approximate duration, and any visible caption/title text.
• If the video can't be directly accessed, ask the user to paste the transcript/description or upload the file — don't stall the whole task on this; proceed with what's available and mark gaps clearly.

Step 2 — Transcription

• Transcribe word-for-word, including filler words, unless user asks for cleaned-up text.
• Timestamp each line (mm:ss or mm:ss–mm:ss ranges).
• Separate narration (VO), on-screen text/captions, and sound effects into distinct labeled lines.

Step 3 — Shot Segmentation

• A "shot" = one continuous visual segment before a hard cut or major transition.
• Number shots sequentially (Shot 1, Shot 2, ...).
• Record start/end timestamp for each shot.

Step 4 — Anchor Image Description

• Anchor image = the single still frame that was likely the source before image-to-video animation was applied.
• Describe it as a standalone image generation prompt: subject, pose, setting, art style, color palette, lighting, composition/framing, mood.
• Write it so it could be pasted directly into an image generator (Midjourney, Nano Banana, DALL-E, etc.) to recreate a similar still.

Step 5 — Motion/Animation Notes

• Identify the type of motion applied to the anchor image: slow zoom-in/out, pan left/right, parallax depth, subtle character movement (blinking, lip-sync, hair/cloth sway), camera shake, particle/light effects.
• Note pacing (slow, medium, fast) and any transition style into the next shot (cut, crossfade, whip pan, match cut).

Step 6 — Assemble

• Combine into the final storyboard table (see format below).
• Include a top-level summary: overall theme, tone, target style, total shot count, estimated total duration.

Step 7 — Flag Gaps

• If a shot's source image is ambiguous (e.g., complex composite, stock footage mixed in, fast cuts), explicitly note: "Anchor image uncertain — appears to blend [X] and [Y]."

Example 1: Input: TikTok link to an AI-animated video narrating the Gospel of Matthew, ~45 seconds, painterly biblical art style. Output: Full transcript of the narration split into 8 timed lines; storyboard of 8 shots each with (a) an anchor image prompt like "Jesus preaching on a hillside, golden hour, Renaissance fresco style, wide shot, warm color palette" and (b) motion notes like "slow parallax pan left, robes gently swaying"; final note flagging Shot 5 as "anchor uncertain — likely two blended images (crowd + sky) due to fast cross-dissolve."

Example 2: Input: A link with no accessible audio, only visual content. Output: Transcription section marked "No spoken narration detected — on-screen text only," listing each caption with timestamp; storyboard and anchor images proceed normally; a note asking user to confirm if there's a voiceover not captured.