# TrueTraceShorts Director Pass + Cinematic Motion Standard

Use this before **every** TrueTraceShorts / AutoShortsBot render, especially pure-AI cinematic scam shorts.

The renderer is not the director. Before rendering, decide emotion, eye direction, pressure, relief, and what must work without sound.

## Hard gate

No render without a Director Pass.

The Director Pass must answer:

1. What does the viewer feel in second 1?
2. What is the open question?
3. What is the dangerous moment?
4. Where should the eye go in the first shot?
5. What must be understandable without sound?
6. Which scene creates pressure?
7. Which scene shows the near-wrong action?
8. Which scene brings relief?
9. Which scene is the strongest screenshot?
10. Which movement is necessary, and which movement would feel cheap?
11. Which audio cues does the video need?
12. Which scene needs a hold instead of movement?

If any answer is vague, do not render yet.

## No fixed six-image rule

Choose image count by dramaturgy:

- 25–30s Red Flag Short: 4–6 images.
- Already Clicked / Recovery: 5–7 images.
- True Scam Story: 5–8 images.
- Quiz: 3–5 images.
- Breakdown: 4–6 images.

One strong AI image may provide multiple virtual shots:

- Wide.
- Medium.
- Close-up.
- Extreme close-up.
- Hold.

Use crops and controlled camera moves to create filmic shots from strong images.

## Shot purpose vocabulary

Every image/virtual shot needs a purpose. Name it as one of:

- Shock.
- Confusion.
- Pressure.
- Near mistake.
- Recognition.
- Safe move.
- Relief.
- Takeaway.

If a shot has no purpose, do not use it.

## Cinematic Motion Engine

Allowed implementation tools:

- FFmpeg for final export, audio, filters, subtitles.
- Python timeline/orchestration.
- Pillow for deterministic frame rendering.
- OpenCV + NumPy for controlled crops, affine transforms, vignette, light/color effects, motion blur if needed.
- MoviePy only if useful as orchestrator.
- PySceneDetect only for analysis of reference / own videos, not copying.

Do not introduce for these scam shorts unless explicitly requested:

- Manim.
- PowerPoint graphics.
- Old UI mockup renderers.
- Wan/I2V.
- Stock video.

## Motion grammar

Allowed:

- controlled push-in.
- X/Z zoom.
- punch-in.
- slow drift.
- parallax-like crop.
- gentle tilt only when motivated.
- short hold.
- focus emphasis.
- subtle vignette.
- subtle alarm color ramp.
- safe-move color reset.

Forbidden:

- artificial shake.
- frame vibration.
- rotation wobble.
- random zooming.
- cheap alarm flash.
- excessive red filter.
- repeated whooshes.
- movement without purpose.

Movement must be eye direction, not effect.

## Hook grammar for scam shorts

0.0–0.4s:
- Person + dangerous screen visible.

0.4–1.4s:
- Push-in toward danger.

1.4–2.2s:
- Close-up / hold on scam text.

2.2–3.0s:
- Emotional hook line.

The first 2 seconds must work without sound.

## Voice direction

Use two voice states:

A. Hook voice:
- more emotional.
- concerned.
- human.
- short.
- not theatrical.

B. Main voice:
- calm.
- clear.
- trustworthy.
- simple.

If Chatterbox allows it:
- render hook as separate TTS clip.
- test max 2 hook variants.
- choose the more natural variant.

Do not make the whole video panicked.

## Audio direction

Every video needs an Audio Plan:

- music bed: low tension, quiet but perceptible.
- SFX 1: notification/ping if appropriate.
- SFX 2: low hit at red flag.
- SFX 3: safe resolve at end.
- duck music under voice.
- no loud music.
- no horror-trailer music.
- no TikTok clown sounds.

Audio must lead the story, not merely fill space. Do not claim “music/SFX were used” solely because waveforms were generated; QA must judge whether the cues are perceptible and narratively timed. If the user says music/SFX are missing, treat that as an audio-direction failure: increase presence slightly, verify loudness/true peak, and re-check that voice remains clearly in front.

Important: final video must not outlast the voiceover unless there is an intentional, explicitly planned visual/audio tail. For normal previews, trim to voice duration + at most ~0.3–0.5s resolve tail. Avoid silent black/idle endings.

After every render, measure both video duration and voice duration. Treat `video_duration - voice_duration > 0.5s` as a failed gate and re-render before delivery. Do not report or ship a preview with an accidental silent tail.

## Automatic QA for every preview

Create and inspect:

### A. First-Second Sheet

Frames:
- 0.0s
- 0.5s
- 1.0s
- 1.5s

### B. Silent Recognition Verdict

Question: Does a silent viewer understand what is dangerous?

### C. Readability Check

Is the scam text readable?

### D. Cheap Effect Check

Were shake, wobble, flash, PowerPoint look, cards, arrows, circles, boxes, labels avoided?

### E. Beat Density Check

Does something meaningful change every 2–4 seconds?

### F. Eye Direction Check

Does the edit guide the eye to the most important object?

### G. Subtitle Logo Safe-Zone Check

TrueTraceShorts channel logo / Shorts UI can cover low subtitles. Position active-word captions about **two subtitle lines higher** than the old baseline. For 1080x1920 renders, keep the caption block around `y=1400–1450` for two-line captions and `y=1450–1500` for one-line captions, not down at ~1530–1580. Verify the frame sheet and preview against the logo-safe area before delivery.

## ReferenceShortAnalysis

- Do not treat generated AI keyframes as safe just because the prompt forbids URLs or data. Inspect them visually. If any generated frame contains `http`, a domain-like fragment, brand/logo, QR, phone, card number, bank-like text, or unintended personal data, regenerate that keyframe before render.
- The safest first frame is often a strong AI still with readable renderer-owned scam text already inside the scene. Captions should normally stay out of the first ~0.5s so they do not compete with silent recognition.
- When creating QA/contact/frame sheets, clamp sample timestamps to `< video_duration` (for example `min(requested, duration - 0.2)`). Otherwise a perfectly valid short can produce a missing/black final QA tile and waste review time.
- If a duration gate fails (`video_duration - voice_duration > 0.5s`), fix the timeline to voice + short resolve tail; never pad with silent visuals to satisfy a nominal target duration.
- For post-private-upload workflow, this project now expects the website companion to be updated immediately and the real Shorts URL to be visible, because the user manually switches the video public right away. Do not hide the link solely because the upload was initially private.

## ReferenceShortAnalysis

Optional tool/function for own published videos or manually provided good references.

Input:
- own published videos.
- later manually supplied good reference videos.

Output:
- cut frequency.
- first-shot duration.
- number of visual beats.
- caption density.
- music/SFX notes.
- average shot length.

Use PySceneDetect only for analysis, not for copying external videos.

## Required report fields after every video

In addition to normal package data, report:

- Director Pass Summary.
- Number of AI images.
- Number of virtual shots.
- Strongest hook shot.
- Strongest safe-move shot.
- Motion Plan.
- Audio Plan.
- First-Second Sheet.
- Silent Recognition Verdict.
- Cheap Effect Verdict.
- What still needs improvement.

## Producer mindset

Do not think: “I have 6 AI images and make a video.”

Think:

- What emotion must the viewer feel in second 1?
- Where must the eye go?
- When does pressure rise?
- When does relief arrive?
- What works without sound?
- Does every movement, cut, and sound have a reason?

If an effect makes the video feel cheaper, remove it. Less, but precise, is better.