Prompt Engineer Skill

Purpose

Transform story and visual concepts into optimized prompts for AI image generation services. This skill provides consistent handling of prompts across different generation modes (photorealistic, concept art, reference sheets) and services (fal.ai models).

Trigger

When generating any visual asset that requires prompt construction.

Inputs Required

• •Visual concept or scene description
• •Target output mode (photorealistic, concept, reference)
• •Style DNA from PROJECT_CONFIG.yaml
• •Character or location data (if applicable)

Outputs Produced

• •Optimized prompt string
• •Negative prompt string
• •Recommended model parameters

Prompt Modes

Mode 1: PHOTOREALISTIC (Production Stills)

Goal: Images that look like frames from an actual film shoot.

Framework: Frame the prompt as a "cinematographer's lookbook" or "location scout's gallery" rather than "concept art" or "illustration."

Required Elements:

1. •

   Camera & Lens Specification

   • •Specific lens: 24mm anamorphic, 50mm f/1.4, 85mm portrait lens
   • •Camera reference: shot on ARRI Alexa, RED camera, Panavision
   • •Depth of field: f/2.8 shallow depth of field, f/8 deep focus
2. •

   Film Stock / Color Science

   • •Film stocks: Kodak Vision3 500T, Fujifilm Eterna, 35mm film grain
   • •Digital looks: ARRI color science, Blackmagic RAW
3. •

   Lighting as Physical Reality

   • •Practical sources: practical light from oil lanterns, window key light
   • •Quality descriptors: hard shadows, diffused daylight, bounce fill
   • •Avoid: "magical glow", "ethereal light", "supernatural illumination"
4. •

   Material Physics Over Vibe Words

   • •DO: damp limestone reflecting amber lantern light, refractive distortions through leaded glass
   • •DON'T: ethereal, mystical, magical, impossible, supernatural, otherworldly
5. •

   Power Phrase

   • •Include: practical set construction - signals physical reality over digital painting

Template:

[SCENE DESCRIPTION]. [CAMERA/LENS]. [LIGHTING QUALITY]. [MATERIAL DETAILS].
Shot on [CAMERA], [FILM STOCK]. [f-STOP]. Practical set construction.

Example:

Caribbean colonial prison cells, damp stone walls with iron bars. 24mm anamorphic
lens, f/2.8. Single shaft of golden afternoon light cuts through dust particles,
illuminating condensation on limestone. Rust oxidation on iron, water stains on
mortar, practical oil lantern visible in frame. Shot on ARRI Alexa, Kodak Vision3
500T color science. Practical set construction.

Mode 2: CONCEPT ART (Development Phase)

Goal: Evocative images for creative exploration and mood setting.

Framework: Painterly, illustrative quality is acceptable. Focus on emotional impact.

Allowed Elements:

• •Atmospheric adjectives: ethereal, haunting, luminous
• •Style references: in the style of, reminiscent of
• •Art medium references: digital painting, concept art, matte painting

Template:

[MOOD] [SCENE DESCRIPTION]. [STYLE REFERENCE]. [COLOR PALETTE].
[ATMOSPHERE DESCRIPTORS]. Cinematic concept art, high detail.

Mode 3: REFERENCE SHEETS (Character/Location Identity)

Goal: Consistent multi-panel layouts for production reference.

Framework: Technical specification document, clinical presentation.

Required Elements:

1. •Grid specification: 2x2 grid, 3x2 layout, 8-panel composite
2. •Background: neutral #2d2d2d background, white studio background
3. •Panel descriptions: Explicit description of each panel position
4. •Consistency anchors: Age, physical attributes, signature elements

Template:

A professional [GRID] [TYPE] reference sheet for [PRODUCTION TYPE]. Subject: [NAME].

[PANEL DESCRIPTIONS - explicit position and content for each]

Style: [STYLE DNA]. Neutral background with thin dividers between panels.

Vocabulary Guide

Photorealistic - AVOID These Words

Photorealistic - USE These Words

Negative Prompt Templates

Photorealistic Mode

digital painting, illustration, concept art, anime style, cartoon,
stylized, CGI render, video game, oversaturated, HDR, neon colors,
fantasy glow, magical effects, lens flare abuse, floating elements,
deformed anatomy, extra limbs, bad proportions, blurry, low quality

Reference Sheet Mode

desaturated, gritty, grimdark, realistic gore, sexualized,
Halloween costume aesthetic, campy, muddy colors, floating heads,
bland lighting, generic fantasy, anime style, cartoon style,
multiple people, crowd scene, deformed, extra limbs, bad anatomy

Model Selection Guide

Process

Step 1: Identify Output Mode

Determine if the request is for:

• •Production stills → PHOTOREALISTIC mode
• •Early development → CONCEPT ART mode
• •Identity/reference → REFERENCE SHEET mode

Step 2: Gather Source Data

• •Read PROJECT_CONFIG.yaml for style DNA
• •Read relevant character/location configs
• •Note any specific visual keywords

Step 3: Build Prompt Structure

Follow the template for the identified mode:

1. •Assemble scene/subject description
2. •Add mode-appropriate technical elements
3. •Apply vocabulary corrections (replace vibe words)
4. •Add style DNA elements
5. •Construct negative prompt

Step 4: Parameter Selection

Based on mode and model:

• •Set aspect ratio (16:9 for locations, square for identity sheets)
• •Set resolution (2K for final, 1K for exploration)
• •Set inference steps (40 for quality, 25 for speed)
• •Set guidance scale (4.5 standard, lower for creativity)

Integration with fal_generate.py

The prompt engineer skill informs these functions:

• •build_identity_prompt() - Reference sheet mode
• •build_hero_prompt() - Photorealistic mode
• •build_location_prompt() - Photorealistic mode with architectural focus

Examples

Input: "Prison cells where the ledger was hidden"

Concept Art Mode Output:

Prompt: Haunting prison cells beneath a Caribbean gallows, shafts of ethereal
golden light piercing the darkness, atmosphere of forgotten confessions and
desperation. Moody cinematic concept art, rich shadows, amber and rust palette.

Negative: cartoon, anime, bright colors, cheerful, modern elements

Photorealistic Mode Output:

Prompt: 18th century Caribbean colonial prison cells, damp limestone walls with
mortar deterioration visible, rust-oxidized iron bars, single shaft of afternoon
sun cutting through dust particles at 45-degree angle. Period-accurate iron
shackles, practical oil lantern mounted on wall bracket. 24mm anamorphic lens,
f/2.8, shallow depth of field on foreground bars. Shot on ARRI Alexa, Kodak
Vision3 500T. Practical set construction, location scout photograph.

Negative: digital painting, illustration, concept art, magical glow,
supernatural lighting, fantasy elements, stylized, CGI render

Quality Checklist

Before finalizing any prompt:

• • Mode-appropriate vocabulary used
• • No conflicting style signals (photorealistic + "concept art")
• • Technical specifications included for photorealistic mode
• • Material physics described, not just mood
• • Negative prompt addresses common failure modes
• • Aspect ratio appropriate for output type

STILL IMAGE: Logical Consistency Rules

When crafting prompts for still images, avoid "logical knots" that confuse the model. Each element must work together physically and cinematically.

Rule 1: Single Time of Day / Lighting State

Problem: Asking for a "transition" (e.g., "golden hour to wrong-blue moonlight") in a single still forces the model to represent change over time.

Result: The model may split the image awkwardly or create muddy orange-teal color casts.

Fix: Describe the result of the transition, not the transition itself. Pick the dominant look.

Rule 2: Lens Logic Must Be Consistent

Problem: Wide-angle lenses (24mm) naturally produce deep focus. Asking for "shallow depth of field" with a wide lens contradicts optical physics.

Cinematography Reality:

Fix: Match lens choice to desired depth of field:

Rule 3: Describe States, Not Motion

Problem: Still images cannot show motion. Describing actions mid-process ("landing", "falling", "running") forces the model to freeze an unstable moment.

Fix: Describe the pose or result of the action:

Rule 4: Conflicting Light Sources

Problem: Multiple light sources with different qualities can create muddy or impossible lighting.

Fix: Establish hierarchy - one KEY light, supporting FILL or ACCENT lights:

Example: Before/After Optimization

Before (logical knots):

Interior of colonial prison at night, golden hour to wrong-blue transition,
young woman landing silently on stone floor having dropped through window,
24mm anamorphic lens, shallow depth of field, long perspective of cell doors

After (logically consistent):

Cinematic wide shot, interior Caribbean colonial prison corridor at night.
A young woman in mid-crouch on weathered stone floor beneath a high barred window.
Practical oil lantern casts flickering amber glow against damp walls.
Wrong-blue teal moonlight streams through bars, creates sharp shadows.
Long perspective of iron cell doors receding into dark misty void.
Shot on ARRI Alexa, 24mm anamorphic lens, deep focus, high contrast,
heavy texture on stone and iron.

Fixes applied:

1. •Single lighting state (moonlight dominant, lantern accent)
2. •Wide lens + deep focus (not shallow DoF)
3. •"Mid-crouch" pose instead of "landing" action

FRAME vs VIDEO PROMPTS (CRITICAL DISTINCTION)

Frame prompts and video prompts serve different purposes. Confusing them causes generation failures.

Frame Prompts (for Nano Banana Pro / image generation)

Purpose: Generate a single, clean moment to use as a video start frame.

Rules:

• •Describe STATIC states, not transitions
• •No temporal language ("then", "as", "shifts to")
• •Single lighting state
• •Single character pose/expression

Video Prompts (for Kling / video generation)

Purpose: Describe motion and transitions that animate FROM the start frame.

Rules:

• •Include motion verbs and camera movement
• •Transitions are allowed ("shifts from X to Y")
• •Describe what CHANGES, not what stays static
• •Must be START FRAME AWARE - continue from visible state

Transitional Language Placement

CRITICAL: If transitional language appears in a frame prompt, it will cause composite images (multiple states rendered in one frame).

VIDEO GENERATION: Kling 3.0 Pro

Overview

Kling 3.0 Pro (fal-ai/kling-video/v3/pro/image-to-video) generates 3-15 second video clips from starting images with strong character/location consistency through its Elements system.

API Schema Requirements

CRITICAL: These formats are enforced by validation - incorrect formats will fail.

Elements Format

Elements provide character/location consistency across the video. BOTH fields are required:

elements = [
    {
        "frontal_image_url": "https://...",      # REQUIRED: Clear frontal view
        "reference_image_urls": ["https://..."]  # REQUIRED: Array of additional angles
    }
]

• •Wrong: {"frontal_image_url": "..."} (missing reference_image_urls)
• •Right: {"frontal_image_url": "...", "reference_image_urls": ["..."]}

Multi-Prompt Format

Multi-prompt enables multiple "cuts" within a single video. Each prompt needs its own duration:

multi_prompt = [
    {"prompt": "Close-up shot...", "duration": "3"},
    {"prompt": "Medium shot...", "duration": "3"},
    {"prompt": "Wide shot...", "duration": "4"}
]

• •Wrong: ["prompt 1", "prompt 2"] (strings, not dicts)
• •Wrong: [{"prompt": "..."}] (missing duration)
• •Right: [{"prompt": "...", "duration": "N"}]
• •Sum Rule: Total of all duration values must equal the video's total duration

Duration Values

• •Always use strings, not integers: "5" not 5
• •Range: "3" to "15" seconds
• •With audio (generate_audio: true): $0.336/second
• •Without audio: $0.224/second

Element Reference Syntax

In prompts, reference uploaded elements with @Element1, @Element2, etc:

"@Element1 walks through the @Element2 market, looking around curiously"

• •@Element1 = First element in array (typically main character)
• •@Element2 = Second element (typically location or secondary character)

Video Prompting Techniques

Start Frame Awareness (CRITICAL)

Video prompts must describe actions that CONTINUE FROM the visible start frame state.

The model SEES the start frame. If the prompt contradicts what's visible, the model will:

• •Try to reconcile the contradiction (creating awkward transitions)
• •Invent intermediate actions (moving to a "new" position)
• •Produce discontinuous motion

Before writing video prompts:

1. •View/understand the start frame (generated shot frame OR extracted last frame)
2. •Note the character's current position/state/expression
3. •Write prompts that continue FROM that state, not TO that state

Motion Verbs (Essential)

Unlike image prompts, video prompts MUST include motion:

Camera Movement Descriptors

Transition Language (Multi-Prompt)

When using multi-prompt, prefix each prompt with "Cut to:" to signal clear scene transitions. This helps the model understand each segment is a distinct shot:

• •Cut 1: "Cut to: Close-up on hands examining..."
• •Cut 2: "Cut to: @Element1 speaks to the merchant, expression shifts from..."
• •Cut 3: "Cut to: Wide shot pulling back as @Element1 turns and walks away..."

Note: The "Cut to:" prefix improves transition clarity and reduces artifacts between segments.

Multi-Cut Scene Template

For a 10-second, 3-cut scene:

multi_prompt = [
    {
        "prompt": "Cut to: Close-up on [DETAIL], [MOOD], [LIGHTING], cinematic shallow depth of field",
        "duration": "3"
    },
    {
        "prompt": "Cut to: @Element1 [ACTION], [EXPRESSION CHANGE], medium close-up, [LIGHTING]",
        "duration": "3"
    },
    {
        "prompt": "Cut to: Wide shot [PULLBACK ACTION] as @Element1 [DEPARTURE ACTION] into @Element2, [ATMOSPHERE], cinematic crane movement",
        "duration": "4"
    }
]

Recommended Parameters

request = {
    "start_image_url": start_url,        # Required: Starting frame
    "prompt": prompt,                      # OR multi_prompt (mutually exclusive)
    "multi_prompt": multi_prompt,
    "elements": elements,                  # Character/location references
    "duration": "10",                      # Total video length
    "aspect_ratio": "16:9",               # "16:9", "9:16", or "1:1"
    "generate_audio": True,               # Native audio generation
    "negative_prompt": "blur, distort, low quality, cartoon, anime, deformed hands",
    # Optional:
    "end_image_url": end_url,             # For transition videos
}

Character Element Best Practices

For maximum character consistency:

1. •Frontal: Use identity sheet or clear frontal portrait
2. •References: Include 2-3 additional angles (action pose, profile, etc.)
3. •Multiple characters: Add as separate elements, reference as @Element1, @Element2

# Good character element setup
elements = [
    {
        "frontal_image_url": upload(identity_sheet),
        "reference_image_urls": [
            upload(hero_shot_1),
            upload(hero_shot_2)
        ]
    }
]

Location Element Best Practices

Locations can also use elements for consistency:

{
    "frontal_image_url": upload(main_location_ref),
    "reference_image_urls": [upload(alternate_angle)]
}

Reference in prompts: "walks through the bustling @Element2 market"

Video Prompt Examples

Single Character Motion:

@Element1 looks up slowly, her eyes catching the golden light, a subtle knowing
smile crosses her face, cinematic slow movement, shallow depth of field

Multi-Character Interaction:

@Element1 and @Element2 exchange a glance, @Element1 reaches out to touch
@Element2's arm, intimate moment, golden hour lighting, cinematic two-shot

Scene Progression (3 cuts):

Cut 1: "Close-up on hands examining strange glowing bottles, curiosity, soft
        ambient light filtering through market stalls"
Cut 2: "@Element1 speaks to an unseen merchant, her expression shifts from
        curiosity to suspicion, medium close-up, warm golden market lighting"
Cut 3: "Wide shot pulling back as @Element1 turns and walks away into the
        bustling @Element2 market, atmospheric haze, golden hour light"

Common Failure Modes

Video Prompts: Descriptive vs Narrative Language

CRITICAL: Video models interpret language literally. Prose or narrative language will be rendered visually.

Rule: Describe only what the CAMERA SEES, not what the CHARACTER FEELS or what the STORY IMPLIES.

Voice vs Elements Trade-off

CRITICAL LIMITATION: Kling 3.0 does not allow voice_ids and elements in the same request.

Workflow for dialogue scenes:

1. •Use a frame showing the character as start_image_url
2. •Include character description in prompts (hair, clothing)
3. •Use voice_ids with <<<voice_N>>> syntax
4. •NO elements array

Workflow for visual consistency scenes:

1. •Use elements with frontal + reference images
2. •Use @ElementN syntax in prompts
3. •NO voice_ids - use native audio generation

Quality Checklist (Video)

Before submitting a Kling 3.0 request:

• • Elements have both frontal_image_url AND reference_image_urls
• • Multi-prompt uses dict format with prompt and duration keys
• • Cut durations sum to total video duration
• • Prompts include motion verbs and camera movement
• • @ElementN references match uploaded element order
• • Duration values are strings ("5" not 5)

VOICE DESIGN: Character Voice Pipeline

Creating consistent character voices for video requires a multi-step pipeline using MiniMax for voice design and Kling for video integration.

Voice Pipeline Overview

MiniMax voice-design → Preview Audio → Kling create-voice → Kling voice_id → Kling video
     ($1.00)              (.mp3)           ($0.035)         (numeric ID)      ($0.392/s)

IMPORTANT: MiniMax voice IDs are NOT compatible with Kling video. You must clone the MiniMax audio using Kling's create-voice endpoint to get a Kling-compatible voice_id.

Step 1: MiniMax Voice Design

Endpoint: fal-ai/minimax/voice-design

Cost: $1.00 per voice + $0.03/1000 chars preview

Use the "Archetype + Modifier" formula for best results:

voice_prompt = """A young female adventurer, 16 years old, with a clear and warm voice
that carries a subtle Caribbean lilt. She speaks with quick wit and easy charm,
her words winding and redirecting like someone used to talking her way out of trouble.
There's a guarded intelligence beneath the friendly surface. Slightly breathless energy,
as if she might need to run at any moment. Confident delivery that masks deeper uncertainty."""

preview_text = "You're asking what I want? That's a big question for someone who just met me."

result = fal_client.subscribe(
    "fal-ai/minimax/voice-design",
    arguments={
        "prompt": voice_prompt,
        "preview_text": preview_text,
    }
)
# Returns: {"custom_voice_id": "ttv-...", "audio": {"url": "https://..."}}

Voice Description Formula:

1. •Archetype: Role-based character (e.g., "young adventurer", "wise mentor", "grumpy sailor")
2. •Physicality: Vocal texture (e.g., "gravelly", "breathy", "crisp", "warm")
3. •Disposition: Emotional baseline (e.g., "guarded", "optimistic", "cynical")
4. •Speech Pattern: Delivery style (e.g., "quick-witted", "slow and deliberate", "winding")

Step 2: Kling Voice Clone

Endpoint: fal-ai/kling-video/create-voice

Cost: $0.035 per voice

Clone the MiniMax preview audio to get a Kling-compatible voice_id:

result = fal_client.subscribe(
    "fal-ai/kling-video/create-voice",
    arguments={
        "voice_url": minimax_audio_url,  # The preview audio from Step 1
    }
)
# Returns: {"voice_id": "848145936509059116"}

Audio Requirements:

• •Duration: 5-30 seconds
• •Clean, noise-free
• •Single voice (no background speakers)
• •Formats: mp3, wav, ogg, m4a, aac

Step 3: Use Voice in Kling Video

Cost: $0.392/second (with voice control)

Reference the Kling voice_id using <<<voice_1>>> syntax directly before the dialogue:

request = {
    "start_image_url": image_url,
    "prompt": "Young woman examines strange bottles, she says <<<voice_1>>> 'I've never seen bottles like these before' expression shifts to suspicion, close-up shot",
    "voice_ids": ["848145936509059116"],  # Kling voice_id from Step 2
    "duration": "5",
    "generate_audio": True,
}

Voice Reference Syntax:

• •<<<voice_1>>> - First voice in voice_ids array
• •<<<voice_2>>> - Second voice (max 2 per generation)

IMPORTANT: Place the voice tag directly before the quoted dialogue, not as a subject:

• •Correct: "She says <<<voice_1>>> 'Hello, stranger.'"
• •Wrong: "<<<voice_1>>> says 'Hello, stranger.'" (voice tag as subject)

MiniMax Voice Prompt Examples

The Clever Young Adventurer (Mars):

A young female adventurer, 16 years old, with a clear and warm voice that carries
a subtle Caribbean lilt. She speaks with quick wit and easy charm, her words winding
and redirecting like someone used to talking her way out of trouble. Guarded intelligence
beneath the friendly surface. Confident delivery that masks deeper uncertainty.

The Gruff Sea Captain:

A middle-aged male with a deep, weathered voice roughened by salt air and shouted orders.
Strong Scottish accent, speaks with gruff authority and a perpetual undertone of impatience.
Words come in short, commanding bursts. Occasional warmth breaks through the harshness.

The Ethereal Presence:

A female voice with an otherworldly quality, clear and resonant as if speaking in a vast
empty space. Calm and measured delivery, slightly detached from emotion. Words carry
weight and significance. Neither warm nor cold, simply present.

The Young Scholar:

A male voice in his late teens, educated and precise in word choice but with youthful
enthusiasm breaking through. Speaks quickly when excited about ideas, slows deliberately
when explaining. Slight breathlessness from constant mental energy.

MiniMax Technical Settings

When generating TTS with MiniMax (if using for non-Kling purposes):

Voice Pipeline Cost Summary

Example 5-second video with voice: $1.00 + $0.035 + ($0.392 × 5) = $2.995

Quality Checklist (Voice)

Before voice pipeline:

• • Character voice description uses Archetype + Modifier formula
• • Preview text captures character's speech patterns
• • MiniMax preview audio saved for reference
• • Kling voice_id stored in project metadata
• • Video prompts use <<<voice_N>>> syntax correctly