Director Prompt Engineer

1. Narrative Arc (Every Director Needs One)

Structure segments with intentional ordering:

• Segment 1 (Hook): Most compelling visual or statement
• Segments 2-3 (Context): Establish setting and subject
• Middle segments (Development): Build complexity and engagement
• Second-to-last (Climax): Most dramatic or revealing moment
• Final segment (Resolution): Conclude with impact

Adapt by use case:

• Documentary: Surprising fact → chronological development → legacy/impact
• Ads: Product tease → problem → solution → social proof → CTA
• Educational: Attention question → concept intro → examples → recap
• Music video: Establish mood → build intensity → peak → denouement

2. Visual Consistency Rules

Include these in every director system prompt:

• Color palette locking: Establish palette in segment 1, repeat same keywords in ALL prompts
• Lighting direction: Choose once (e.g., "dramatic side-lighting from the left"), repeat everywhere
• Character anchors: 15-20 word appearance description, pasted verbatim in every prompt featuring that character
• Environment anchors: 10-word location description, pasted verbatim for recurring settings
• Style keyword: Include {{Style}} in EVERY image and video prompt
• Negative constraints: "No text, labels, titles, or watermarks" in every visual prompt

3. Camera Movement Instructions

Always instruct the director to specify camera movement. Without them, video models produce static or random motion:

• Dolly: Forward/backward — approaching subjects
• Pan: Horizontal rotation — revealing wide scenes
• Tilt: Vertical rotation — tall subjects or reveals
• Crane: Vertical movement — establishing shots
• Tracking: Following subject — action sequences
• Orbit: Circular around subject — dramatic reveals

4. TTS-Friendly Narration

Narration will be read by text-to-speech engines. Instruct the director to:

• Write ONLY spoken words (no stage directions)
• Spell out numbers: "nineteen forty-five" not "1945"
• Spell out abbreviations: "United States" not "US"
• Use em-dashes for dramatic pauses
• Vary sentence length: short-medium-long-short
• Match energy to content: action → punchy; reflective → flowing

5. Timing Enforcement

The most common failure is narration exceeding segment duration.

Formula: 2 words per second maximum.

Include calibration examples in the system prompt so the LLM can self-check.

6. Conditional Field Handling

When a JSON schema requires fields that are only relevant for some segment types:

• Use empty string "" for inapplicable fields
• Never use "N/A" or placeholder text — these get sent to downstream AI models
• Add explicit instructions: "If NarrationType is ImageNarration, set TalkingHeadText to empty string"

7. [cut] Scene Syntax

For videos with internal scene transitions:

Camera transition description between scenes
First scene camera and action (inherits from input start image)
[cut] Second scene with new angle and camera movement
[cut] Third scene with different perspective

Number of [cut] markers = CutScenesPerSegment - 1 (the input image defines scene 1).

File Structure

A prompt producer consists of three files:

producer.yaml