Video Prompt Engineering

# Video Prompt Engineering Reference

> Load this when constructing video prompts or when the user asks about
> video prompting techniques. Do NOT load at startup.
>
> Based on Google's 2026 "Ultimate Prompting Guide for VEO 3.1."

## The 5-Part Video Prompt Framework

Write as natural narrative prose -- NEVER as keyword lists. Each video prompt should include all 5 parts.

**Target length: 100–200 words.** Shorter prompts underspecify the scene
and produce generic results. Longer prompts (>300 words) tend to
contradict themselves as the model tries to satisfy conflicting details.
**150 words is the sweet spot** — enough to cover all 5 framework parts
without overloading. The VEO 3.1 API hard limit is 1,024 tokens (~4,096
characters); `video_generate.py` warns at ~950 tokens and errors at
~1,125.

### Part 1 -- CAMERA

Shot type and movement. Use professional cinematography language.

**Shot types:** establishing wide, medium shot, close-up, extreme close-up, over-the-shoulder, bird's-eye, worm's-eye, Dutch angle, POV

**Camera movements:**
- **Dolly:** Camera physically moves closer/farther ("slow dolly forward")
- **Tracking:** Camera follows moving subject laterally ("tracking shot alongside")
- **Pan:** Horizontal sweep ("slow pan left across skyline")
- **Tilt:** Vertical sweep ("tilt up from feet to face")
- **Crane:** Vertical arc with reveal ("crane shot rising above the crowd")
- **Zoom:** Lens focal change ("slow zoom into the subject's eyes")
- **Static:** No movement ("locked-off tripod shot")
- **Handheld:** Natural camera shake ("handheld following the action")

**Performance notes:** Zooms have higher success rates than dollies. Lateral tracking works 68% of the time. Start simple, iterate.

**Good:** "Slow dolly forward through the glass door, transitioning from exterior to interior in a single continuous take"
**Bad:** "camera moves"

**Lens focal length** (VEO responds accurately to specific mm values; use them):

| Focal length | Effect | Use case |
|---|---|---|
| 16mm | Expands space, exaggerates depth | Establishing, environments, dramatic interiors |
| 24–35mm | Natural perspective, close to human vision | Standard coverage, documentary, medium shots |
| 50mm | Slight compression, flattering proportions | Portrait, conversation, product closeups |
| 85mm | Heavy background compression, creamy bokeh | Portrait beauty, isolating subject from environment |
| 100mm+ macro | Extreme close-up with very shallow DOF | Product details, textures, pour shots |

Lenses control depth perception, not distance — specify them
explicitly, e.g. `"85mm f/1.4, shallow depth of field"`.

### Part 2 -- SUBJECT

Who or what is in frame. Same specificity as image prompts. Include motion state.

**Good:** "A woman in her 30s with auburn hair in a loose bun, wearing a cream linen apron over a navy henley, flour-dusted hands shaping sourdough"
**Bad:** "a baker"

### Part 3 -- ACTION

What happens during the clip. Must be completable within 4-8 seconds. One dominant action per clip.

**Good:** "She lifts the shaped dough and places it gently into a banneton basket, pressing the surface smooth with her palm"
**Bad:** "she bakes bread" (too broad for 8 seconds)

**Timing rules:**
- 4s: Single gesture, reaction, or beat
- 6s: One action with follow-through
- 8s: One complete micro-story (setup + action + result)

### Part 4 -- SETTING

Location, atmosphere, time of day. Include environmental audio cues.

**Good:** "Inside a warm artisan bakery at dawn, golden light streaming through flour-dusted windows, wooden shelves stacked with cooling loaves"
**Bad:** "bakery, morning"

### Part 5 -- STYLE + AUDIO

Film style, lighting, color grade, AND audio elements. Audio is unique to video.

**Style anchors:** "Shot on ARRI Alexa 65 with Cooke S7/i lenses, warm Kodak Vision3 250D color science, shallow depth of field. Documentary-style like a Chef's Table episode."

**Audio elements (include at least one):**
- **Dialogue:** `She says, "Almost there."` (in quotes)
- **SFX:** `SFX: soft thud of dough hitting the counter, flour puffing into the air`
- **Ambient:** `Quiet hum of the oven, distant birdsong through an open window`
- **Music:** `Gentle acoustic guitar melody in the background`

**Dialogue timing (one breath per clip):**

- **4 s clip:** 3–5 words max (e.g., `A woman says, "I'm ready."`)
- **6 s clip:** 6–10 words (e.g., `He whispers, "Let's go before they see us."`)
- **8 s clip:** 10–15 words — one full breath of natural speech

Too much dialogue causes unnaturally rapid speech; too little leaves
VEO filling the silence with gibberish murmuring in clips that contain
a visible speaker. Never leave a visible speaker silent.

## Timestamp Prompting: Multi-Shot in a Single Clip

VEO 3.1 accepts a timestamp syntax that directs a micro-sequence within
a single 4/6/8-second generation. A 30-second sequence that would
otherwise need 4 separate VEO calls can sometimes be packed into 2
calls with 4 sub-shots each — a ~50% cost reduction.