Prompt Engineering & Consistency Formula

Achieve 99% shot-to-shot consistency on Google Flow. Master the Subject · Scene · Style pre-production framework, keyframe-to-video motion, and deploy our universal AI director prompt to storyboard any short film with precision.

1 The Core Batch Rule: Blank Line Separation

Whether you upload a .txt file or paste a list directly into the side panel, every prompt must be separated by at least one blank empty line.

prompts.txt
Ultra-realistic cinematic shot of a futuristic cyberpunk racer leaning into a high-speed neon turn in Tokyo rain, volumetric reflections, 8k, shot on 35mm lens.

Close-up emotive portrait of the driver removing her helmet, beads of rain on her visor, glowing HUD interface reflections across her eyes, dramatic studio lighting.

Wide establishing angle of the futuristic pit-stop garage, mechanics working on the hover-car chassis, sparks falling, cinematic atmosphere.
How FlowDirector processes this: The panel automatically counts and separates this into 3 distinct jobs in your queue, executing each sequentially without you needing to click or wait.

2 Dynamic Tagging: Labeling Outputs (#name)

Keep consistent characters, environments, or vehicles across a sequence. Put #name anywhere in a prompt to automatically assign that label to the resulting image when Google finishes rendering it.

Prompt Example
#cyber_pilot Character turnaround sheet of an athletic female cyberpunk pilot with silver cropped hair, matte obsidian flight suit with teal luminous circuitry, photorealistic, neutral expression, solid white background.
Automatic & Clean: FlowDirector captures the completed image and stores it in your session reference library under cyber_pilot. The #cyber_pilot tag is cleanly stripped before dispatching to Google Flow so it never pollutes render generation.

3 Reusing Characters & Assets (@tag)

Once an asset has been tagged with #name (or uploaded into your reference library), inject it directly into subsequent shot prompts using @tag.

Downstream Prompt
@cyber_pilot sitting inside the cramped cockpit of a high-speed orbital interceptor, looking out at starry nebula, tense atmosphere.

FlowDirector automatically binds the rendered pilot image from Prompt 1 as the reference image, preserving facial bone structure, hairstyle, and wardrobe across every subsequent shot.

4 The 99% Consistency Formula: Subject · Scene · Style

In Google Flow, reference images are natively organized into three distinct slots under the Reference Images inspector:

Reference Images Subject · Scene · Style
SUBJECT who / what (up to 100) Library
Click to add character turnaround (solid white bg)
SCENE where it happens (up to 100) Library
Click to add empty environment plate (zero humans)
STYLE visual style / mood (up to 100) Library
Click to add master lighting & film stock anchor

When creators fail to achieve consistency, 95% of the time it is because they bake the environment into the character sheet or fail to isolate styles. Follow this strict pre-production workflow:

👤

SUBJECT (who / what)

Locks facial bone structure, skin texture, hair, wardrobe, and physical anatomy via character model sheets on a solid white background.

🏞️

SCENE (where it happens)

Locks spatial geography, architecture, horizon line, environmental weather, and lighting bounce using empty plates with zero humans.

🎨

STYLE (visual style / mood)

Locks camera lens optics (35mm anamorphic), film grain stock (Kodak Vision3), color grade, and global lighting temperature.

The Golden Rule: Character Model Sheets MUST Be on a Plain Solid White Background Never generate character concept sheets in the rain, inside a dark forest, or sitting in mud. Google Flow's diffusion models treat environmental rain streaks and mud splatters as permanent facial and clothing features! Generate on a clean, solid white studio background with flat 3-point lighting to prevent visual drift.

The Master Prompt Assembly Formula: First analyze the story, divide into scenes, define the 3 reference pillars, and then chain your references in this precise order:

Formula
@subject @scene @style [Subject Action, Precise Camera Angle, Environmental Interaction, Atmospheric Physics]

5 Video Motion Modes: @start & @end Keyframing

In Google Flow's Video mode, FlowDirector gives you surgical control over clip transitions and motion dynamics using keyframe anchors:

🎬 Anchor First Frame @start:name

Locks the opening frame of the video to any approved still keyframe. Google Veo animates from that exact visual state with zero facial distortion.

@start:shot1_still The pilot flips a toggle switch overhead as cockpit instruments glow to life, subtle camera vibration.

⚡ Start + End Interpolation @start & @end

Interpolates realistic camera motion and physical transformation smoothly between two distinct keyframe stills.

@start:door_closed @end:runway_open The heavy blast door slides upward revealing the brightly lit desert runway, high wind.

6 The AI Pre-Production Meta-Prompt

Copy and paste this universal director prompt into ChatGPT, Claude, Gemini, or DeepSeek alongside your story. It instructs the AI to automatically analyze your narrative, divide it into scenes, define the 3 Reference Pillars, and output a complete FlowDirector batch list.

System Prompt
You are an elite AI Film Director and Pre-Production Specialist for Google Flow (Veo). Your mission is to take the user's story or script and convert it into a 99% visually consistent, batch-ready storyboard for the FlowDirector automator.

CRITICAL ARCHITECTURE RULES:
1. STEP-BY-STEP PRE-PRODUCTION SEQUENCE:
   - Step A: Analyze the narrative and segment into sequential scenes.
   - Step B: Identify and isolate characters (SUBJECT), locations (SCENE), and cinematic aesthetic (STYLE).
   - Step C: Generate Phase 1 isolated reference assets with tags (#name).
   - Step D: Assemble Phase 2 shot keyframes referencing (@subject @scene @style).
   - Step E: Animate Phase 3 video motion using @start:shot_still.

2. THE 3-PILLAR REFERENCE FOUNDATION (Google Flow Reference Images):
   - SUBJECT (who / what): Character sheets MUST be generated on a PLAIN SOLID WHITE BACKGROUND with flat, neutral 3-point studio lighting and a neutral expression. Never include rain, mud, darkness, or environmental debris in subject sheets.
   - SCENE (where it happens): Environmental plates MUST be EMPTY plates (explicitly specify "zero people, no characters, empty establishing shot").
   - STYLE (visual style / mood): A single unified visual anchor (e.g., 35mm anamorphic lens, Kodak 5219 film stock, organic halation, cinematic teal & amber grade).

3. VERBOSITY CONSTRAINT:
   - For all long-form, narrative, or descriptive attributes, reduce detail and verbosity by 50% compared to standard verbose responses. Prioritize punchy, cinematic keyword density over fluff.

4. FLOWDIRECTOR SYNTAX & BATCHING:
   - Output every single prompt separated by at least ONE BLANK EMPTY LINE.
   - Use #name to tag new reference outputs (e.g. #char_anu, #env_forest, #style_cinematic).
   - Use @name to reference established assets in keyframe still prompts.
   - Use @start:shot_still in Video Mode to animate the approved still plate without facial deformation.

OUTPUT FORMAT:
----------------------------------------
### PHASE 1: 3-PILLAR REFERENCE ASSETS (IMAGE MODE)
1. SUBJECTS (Solid White Background):
#char_name1 Full-body character turnaround model sheet of [description], front and side view, neutral expression, crisp studio 3-point lighting, plain solid white background, 8k photorealistic.

2. SCENES (Empty Plates):
#env_scene1 Empty establishing shot of [location], zero people, completely empty plate, volumetric lighting, atmospheric depth, cinematic 8k.

3. STYLE ANCHOR:
#style_anchor Master visual style key: [Cinematic lighting, camera lens, color grading, film stock, grain texture], 8k.

----------------------------------------
### PHASE 2: SHOT KEYFRAME STILLS (IMAGE MODE)
#shot1_still @char_name1 @env_scene1 @style_anchor Establishing wide shot: [Character pose and position in environment], [exact lighting], [shot composition].

----------------------------------------
### PHASE 3: SHOT-BY-SHOT ANIMATION (VIDEO MODE via @start)
@start:shot1_still Slow cinematic tracking camera: [Character performs action], [camera dynamics], [environmental physics]. Dialogue: "[Spoken line]". Audio: [Sound FX].
----------------------------------------

Now, analyze the following story and generate the complete pre-production package:
[INSERT YOUR STORY / SCRIPT HERE]

7 Production Case Study: "The Guardian in the Rain"

Here is the exact prompt package for an emotional sci-fi short. Notice the 3-step progression: Pillars (clean assets) → Keyframe Stills (approved framing) → Video Motion (animating approved stills).

Step 1: Reference Asset Generation (Image Mode)

Generate clean, isolated assets first. Notice the strict adherence to "solid white background" for characters and "zero humans" for environments.

Phase 1 Prompts
#char_anu Full-body character turnaround model sheet of a 26-year-old South Asian female archaeologist, short messy black bob haircut, expressive brown eyes, wearing a weathered olive utility jacket. Neutral expression, plain solid white background, flat neutral studio lighting, 8k.

#env_forest Empty establishing wide shot of a dense ancient temperate rainforest, colossal moss-draped cedar trees, muddy forest floor, heavy grey mist, pouring rain, zero humans, completely empty environment.

#style_cinematic Master cinematic film aesthetic, Panavision 35mm anamorphic lens, Kodak Vision3 500T film stock, gentle organic film grain, shallow depth of field, natural volumetric lighting, teal and warm amber color palette, 8k.

Step 2: Keyframe Stills Assembly (Image Mode)

Combine your clean character, empty scene plate, and style anchor using @tag references to generate perfect continuity stills.

Phase 2 Prompts
#shot1_still @char_anu @env_forest @style_cinematic Wide cinematic establishing angle: Anu stands among massive wet tree trunks, rain soaking through her olive jacket, holding an illuminated scanner looking down at ancient glowing glyphs.

#shot4_still @char_anu @env_forest @style_cinematic Intimate low-angle close-up: Anu kneeling in the wet forest mud, gently extending her bare hand toward a trembling neon luminescent bioluminescent forest fawn, warm amber fill light reflecting in her damp eyes.

Step 3: Shot-by-Shot Video Animation (Video Mode)

Animate the approved keyframe stills using Image-to-Video for 99% facial & physical stability via the @start: command.

Phase 3 Prompts
@start:shot1_still Tracking camera follows Anu as she navigates through torrential rain, water dripping from her hood as she checks radar telemetry. Audio: Heavy torrential downpour, muddy footsteps.

@start:shot4_still Slow smooth crane shot down: Anu lowers her stance with extreme care, whispering gently to calm the creature. Dialogue: "Hey little one... I'm not going to hurt you." Audio: Gentle whisper, ambient drizzle.

8 Professional Tips for Smooth Batches

⚡

Speed Selection

Use Fast for quick batches. If Google Flow shows rate-limiting notices, switch to Balanced (2 concurrent) or Slow (1 at a time).

🧼

In-Memory Watermark Cleaner

Enable the Watermark Cleaner in settings to automatically output pristine clean 4K videos and images with no manual cropping.

📁

Auto-Download & Folder

Keep "Auto-download" checked to organize completed renders into tidy run-folders in your Downloads directory automatically.

🔄

3-Account Cycling

FlowDirector supports cycling through up to 3 Google accounts to triple your daily generation credits without interruption.

💬 Join the AI Talks Telugu Community
Get free automation prompts, storyboard workflows, and direct creator support on Telegram.
Join Telegram (@aitalkstelugu) ↗