Prompt Bible V3 Core: Quality AI slop guide for Seedream & Seedance
Just slap this into your favorite LLM and tell it to use this as guide/framework/rule set for creating Seedream/Seedance prompts.
# Prompt Bible V3 Core
## Part 1 – Foundations: Intent Over Words
### Seedream / Seedance Edition
---
# What is Prompt Bible V3?
Prompt Bible V3 is a methodology for translating human intent into prompts that AI image and video generation models can interpret as accurately as possible.
The objective is not to create longer prompts.
The objective is not to use more descriptive language.
The objective is not to maximize prompt complexity.
The objective is to minimize the distance between:
> Human Intent → Model Interpretation → Generated Result
Everything in V3 exists to reduce that gap.
---
# Core Principle — Intent Preservation
The most important concept in Prompt Bible V3 is Intent Preservation.
People naturally think in ideas.
Models interpret structured language.
A prompt is the translation layer between those two systems.
Good prompting is therefore not about writing more.
It is about preserving meaning while translating it into a form the model understands.
When a prompt fails, the cause is often not missing detail.
Instead, somewhere between the user's imagination and the model's interpretation, the intended meaning changed.
The role of V3 is to prevent that change whenever possible.
---
# Why V3 Exists
Most prompting advice focuses on describing the final image.
V3 instead focuses on how the model builds its internal interpretation.
Those are not the same thing.
Consider:
> "A tiny fairy."
A human instantly imagines something much smaller than a person.
A model may simply generate a small woman standing in a forest.
The words are correct.
The interpretation is not.
V3 addresses this by asking a different question:
What information does the model need in order to arrive at the same interpretation as the user?
Sometimes the answer is more detail.
Sometimes it is fewer words.
Sometimes it is describing relationships instead of properties.
The goal is always the same:
Preserve the intended interpretation.
---
# Core Principle — Prompts Are Hierarchies
Prompts are not collections of independent descriptors.
They are hierarchies of meaning.
Some information is foundational.
Some information is decorative.
The model cannot assign equal importance to every token.
Likewise, humans do not imagine every aspect of a scene with equal importance.
V3 therefore organizes prompts according to importance rather than category.
The recommended hierarchy is:
1. Identity
2. Defining traits
3. Action
4. Relationships
5. Environment
6. Style
This order reflects how interpretation naturally develops.
First the model should understand what it is looking at.
Then what makes it unique.
Then what it is doing.
Only afterwards should secondary information refine the scene.
---
# Why Hierarchy Matters
Imagine the prompt begins with:
> "A dramatic cinematic sunset with volumetric lighting..."
Only later does it reveal:
> "...featuring a tiny mechanical hummingbird."
The model has already spent attention establishing atmosphere before discovering the subject.
Reversing the order gives the subject priority.
Instead:
> "A tiny mechanical hummingbird hovering beside a blooming flower, illuminated by dramatic cinematic sunset lighting..."
The scene remains the same.
The interpretation becomes much more stable.
Hierarchy determines emphasis.
---
# Core Principle — Every Detail Has a Cost
Models have finite attention.
Every descriptor consumes part of that attention budget.
This does not mean prompts should always be short.
It means every sentence should justify its existence.
High-value information includes:
* Subject identity
* Defining characteristics
* Important actions
* Relationships
* Critical objects
Lower-value information often includes:
* Repeated adjectives
* Decorative synonyms
* Redundant style modifiers
* Multiple descriptions expressing the same concept
Removing redundant wording often improves clarity more than adding new wording.
V3 therefore values information density over word count.
---
# Core Principle — Interpretation Before Literal Wording
Models do not understand language the way humans do.
They construct an interpretation.
That interpretation ultimately determines the generated result.
For this reason, V3 evaluates prompts by asking:
> "What interpretation is the model likely to construct?"
rather than:
> "Are these words technically correct?"
A prompt may be grammatically perfect while still encouraging the wrong interpretation.
Likewise, a prompt may omit unnecessary wording while producing exactly the intended result.
The measure of quality is interpretation fidelity, not literary quality.
---
# Principles and Rules
Prompt Bible V3 separates foundational ideas from implementation guidance.
## Principles
Principles explain why something works.
They should remain stable even as models improve.
Examples include:
* Preserve intent over wording.
* Prompts form hierarchies of meaning.
* Interpretation matters more than literal phrasing.
* Every detail has an attention cost.
Principles are expected to remain broadly applicable across future model generations.
---
## Rules
Rules describe practical ways to apply those principles.
Examples include:
* Front-load subject identity.
* Reinforce defining characteristics.
* Remove contradictory descriptions.
* Eliminate redundant wording.
* Anchor important relationships explicitly.
Rules may evolve as models change.
Their purpose is to serve the underlying principles, not replace them.
---
# Practical Example
Suppose the intended concept is:
> "A fairy only ten centimeters tall."
A literal prompt might simply state the height.
However, a model may struggle to visualize that scale consistently.
Applying the principle of Interpretation Before Literal Wording leads to a different solution:
> "A tiny fairy standing on the rim of a ceramic coffee mug, her feet no larger than the handle's width."
The second prompt communicates the same idea through visible relationships.
The model no longer has to infer what ten centimeters looks like.
It can see it.
The wording changed.
The intent remained the same.
---
# Part 1 Summary
Prompt Bible V3 is built on four foundational principles:
* Preserve intent over wording.
* Prompts are hierarchies of meaning.
* Every detail consumes attention.
* Optimize for the model's interpretation, not the prompt's prose.
Every rule introduced in later chapters exists to support one or more of these principles.
If a future rule ever contradicts these foundations, the principle takes precedence.
---
# Prompt Bible V3 Core
## Part 2 – Identity, Archetypes & Visual Reasoning
### Seedream / Seedance Edition
---
# Introduction
Part 1 established how V3 thinks.
This chapter explains how V3 describes.
Most generation failures are not caused by a lack of descriptive language.
They occur because the model constructs a different mental interpretation than the user intended.
The purpose of this chapter is to reduce that gap by strengthening how subjects are introduced, how archetypes are preserved, and how abstract ideas are translated into concrete visual evidence.
---
# Core Principle — Identity Before Description
When humans recognize something, they first determine what it is, then begin filling in details.
Image and video models appear to behave similarly.
A subject with an uncertain identity will often remain unstable, regardless of how many descriptive details follow.
For this reason, V3 establishes identity before refinement.
Instead of asking:
> "How should this character look?"
First ask:
> "What must the model recognize this subject as?"
Everything else builds upon that foundation.
---
# Rule — Establish Identity Early
Whenever possible, establish the subject before describing secondary characteristics.
Instead of:
> A beautiful cinematic sunset with dramatic clouds surrounding a warrior...
Prefer:
> A battle-worn warrior standing atop a ruined fortress beneath a dramatic cinematic sunset.
The second version gives the model an anchor before introducing supporting information.
This reduces the chance that atmosphere competes with subject identity.
---
# Core Principle — Identity Is More Than Appearance
Identity is not simply a collection of physical traits.
It is the combination of characteristics that makes a subject recognizable.
For recurring characters, identity is reinforced through multiple complementary anchors.
Common identity anchors include:
* Silhouette
* Proportions
* Distinctive physical features
* Color palette
* Clothing
* Typical posture
* Characteristic expressions
* Behavior
* Relationships with other characters or objects
No single anchor must carry the entire burden.
Instead, several compatible anchors work together to strengthen recognition.
---
# Core Principle — Archetypes Compress Meaning
Humans frequently communicate using archetypes.
Words such as:
* Knight
* Fairy
* Dragon
* Witch
* Pirate
* Android
carry far more information than their dictionary definitions.
They imply shape, behavior, proportions, clothing, cultural expectations and visual language.
These words function as compressed visual concepts.
Replacing them with literal descriptions often removes information rather than adding it.
---
# Rule — Preserve Useful Archetypes
If an archetype accurately represents the intended concept, preserve it.
Do not replace:
> fairy
with
> small winged woman
unless the intention is to deliberately avoid the fairy archetype.
Likewise,
> pirate
contains more useful visual information than
> person wearing historical sailing clothes.
Archetypes provide the model with an efficient starting point.
The role of the prompt is to refine them, not erase them.
---
# Core Principle — Archetype Gravity
Observed Behavior (Seedream / Seedance)
Models naturally drift toward their strongest learned interpretations of common concepts.
This tendency can be thought of as Archetype Gravity.
For example:
A prompt requesting an "angel" may naturally pull toward:
* white robes
* feathered wings
* halos
* serene expressions
even if only some of those traits were intended.
Likewise, requesting a "robot" may encourage metallic humanoids even when the intended design is mechanical but organic in silhouette.
Archetypes are powerful because they compress meaning.
They are also powerful because they attract familiar interpretations.
---
# Rule — Reinforce What Makes Your Version Different
When modifying a familiar archetype, clearly establish the features that distinguish it from the default interpretation.
Instead of merely writing:
> futuristic knight
Explain what makes that concept unique.
For example:
> A knight wearing polished medieval plate armor integrated with glowing holographic circuitry.
The archetype remains intact.
Only the interpretation changes.
---
# Core Principle — Visual Attractor Lock
Some visual traits dominate perception.
Once established, they strongly influence how the rest of the subject is interpreted.
Examples include:
* unusual proportions
* distinctive silhouettes
* species
* costumes
* iconic accessories
These traits should be introduced early and reinforced consistently.
Failing to establish them clearly allows the model to drift toward more common alternatives.
---
# Core Principle — Relationships Are Visual Information
Humans rarely perceive objects in isolation.
We understand scenes through relationships.
Examples include:
* standing beside
* towering over
* holding
* looking toward
* emerging from
* sitting beneath
Relationships communicate context that isolated descriptions cannot.
---
# Rule — Describe Relationships Explicitly
Instead of listing independent objects:
> A knight. A horse. A castle.
Describe how they relate:
> A knight mounted on horseback approaching the gates of an ancient stone castle.
Relationships naturally organize scenes.
---
# Core Principle — Scale Is Relative
Measurements communicate little unless they are visually meaningful.
A statement such as:
> "The creature is two meters tall."
requires the model to internally convert that number into a visual representation.
Environmental relationships reduce that burden.
---
# Rule — Anchor Scale Through Comparison
Whenever scale is important, compare the subject against recognizable references.
Examples:
Instead of:
> A ten-centimeter fairy.
Prefer:
> A fairy standing comfortably inside a teacup.
Instead of:
> A fifty-meter dragon.
Prefer:
> A dragon whose wings cast shadows across an entire village.
Relationships communicate scale more reliably than numbers alone.
---
# Core Principle — Translate Ideas Into Visible Evidence
Models generate images.
Therefore, prompts should describe things the model can actually render.
Many concepts are abstract.
Examples include:
* wisdom
* danger
* exhaustion
* elegance
* ancient
* sacred
These cannot be drawn directly.
They must be translated into visible evidence.
---
# Rule — Show, Don't Merely Tell
Instead of:
> An exhausted traveler.
Describe:
> An exhausted traveler with slumped shoulders, dust-covered clothing, and heavy eyes leaning against a weathered walking staff.
Instead of:
> An ancient sword.
Describe:
> A sword with worn leather wrapping, chipped edges, faded engravings, and patches of darkened steel from centuries of use.
The model no longer needs to invent what "ancient" or "exhausted" might look like.
It is given observable evidence.
---
# Observed Behavior — Visual Evidence Improves Stability
Seedream and Seedance consistently appear to produce more stable interpretations when abstract concepts are expressed through visible consequences rather than isolated descriptive labels.
This does not mean descriptive words should never be used.
Rather, they become significantly stronger when supported by concrete visual evidence.
For example:
> frightened
becomes more reliable when accompanied by:
* widened eyes
* tense posture
* cautious movement
* protective body language
The emotion becomes visible rather than implied.
---
# Part 2 Summary
This chapter introduced the visual reasoning principles that underpin Prompt Bible V3:
* Establish identity before refinement.
* Preserve useful archetypes.
* Reinforce what distinguishes your interpretation.
* Use relationships to organize scenes.
* Anchor scale through comparison.
* Translate abstract ideas into visible evidence.
Together, these principles help ensure that the model constructs the same mental representation the user intended before additional details are introduced.
---
# Prompt Bible V3 Core
## Part 3 – Motion, Physics & Scene Construction
### Seedream / Seedance Edition
---
# Introduction
The previous chapters focused on what the model should imagine.
This chapter focuses on what should happen.
One of the biggest differences between an image that feels posed and one that feels alive is whether the scene obeys believable physical relationships.
People rarely imagine a frozen collection of objects.
They imagine moments.
Moments have momentum.
They have causes.
They have consequences.
Prompt Bible V3 therefore treats actions as physical events rather than isolated verbs.
---
# Core Principle — Every Action Creates Consequences
Nothing moves in isolation.
Every action changes something.
If someone runs:
* clothing moves
* hair shifts
* feet interact with the ground
* dust may be disturbed
* posture changes
If a sword swings:
* the arms rotate
* clothing follows the motion
* the body shifts weight
* nearby objects may react
A believable scene is built from connected consequences.
---
# Rule — Describe Actions as Chains
Instead of writing:
> A knight swinging a sword.
Think in sequence.
Beginning.
Action.
Result.
For example:
> A knight lunges forward, swinging a longsword in a wide horizontal arc as loose fabric trails behind the motion and scattered leaves lift from the ground.
The action now has momentum.
---
# Core Principle — State → Action → Consequence
Most dynamic scenes naturally follow three stages.
State
Where everything begins.
Action
What changes.
Consequence
How the world responds.
This sequence helps maintain logical continuity.
It also provides the model with enough context to build believable motion.
---
# Example
Instead of:
> A wizard casting magic.
Describe:
State
A wizard raises one hand toward the sky.
Action
Blue arcs of magical energy gather around the fingertips.
Consequence
Nearby grass bends outward under the magical force as glowing particles spiral through the air.
The scene now evolves rather than merely existing.
---
# Core Principle — Motion Propagates
Movement spreads through connected objects.
When one object moves, nearby objects often respond.
Examples include:
Walking:
* clothing shifts
* backpack bounces
* hair follows movement
Strong wind:
* trees bend
* loose fabric flutters
* dust moves
* water ripples
Explosion:
* debris flies
* smoke expands
* nearby characters react
* lighting changes
Motion rarely affects only one object.
---
# Rule — Think Beyond the Subject
Ask:
"What else would react?"
These secondary effects dramatically increase realism.
Often, they require only a few words.
---
# Core Principle — Materials Behave Differently
Not every object responds to motion in the same way.
Understanding material behavior strengthens visual consistency.
Examples:
Hair
Flows.
Cloth
Folds.
Metal
Maintains structure.
Glass
Reflects and fractures.
Smoke
Expands and disperses.
Water
Splashes, ripples and reflects.
Fire
Flickers, stretches and changes shape.
The prompt should respect these differences whenever material interaction is important.
---
# Rule — Describe Material Behavior, Not Just Materials
Instead of:
> Flowing cape.
Describe:
> A heavy cape billowing behind the character as they sprint through strong wind.
Instead of:
> Smoke.
Describe:
> Thick smoke slowly curling upward through broken roof beams.
Behavior communicates material properties naturally.
---
# Core Principle — The Environment Is Part of the Action
Scenes become more believable when the environment participates.
Examples:
Rain
Creates puddles.
Darkens clothing.
Generates reflections.
Snow
Accumulates.
Compresses under footsteps.
Falls from tree branches.
Desert wind
Carries sand.
Reduces visibility.
Shapes dunes.
The environment should react when appropriate.
---
# Rule — Avoid Floating Subjects
Characters should interact with their surroundings.
Examples include:
Standing in water.
Leaning against walls.
Leaving footprints.
Casting shadows.
Holding objects.
Breaking branches.
Compressing grass.
Every interaction reinforces presence.
---
# Core Principle — Camera Commitment
A camera can only occupy one place at a time.
The prompt should therefore maintain a consistent visual perspective unless a deliberate transition is intended.
Mixing incompatible viewpoints introduces ambiguity.
---
# Rule — Commit to a Viewpoint
Examples include:
Eye level.
Low angle.
High angle.
Close-up.
Wide establishing shot.
Over-the-shoulder.
These perspectives each emphasize different information.
Switching between them unintentionally can weaken composition.
---
# Observed Behavior — Clear Camera Language Improves Consistency
Seedream and Seedance generally produce more coherent compositions when the intended viewpoint is established early.
This is especially important for video generation, where inconsistent perspective can produce noticeable continuity errors between shots.
---
# Core Principle — Scenes Should Feel Lived In
Many prompts describe objects.
Fewer describe evidence that something has happened.
Signs of interaction often make scenes feel more convincing than adding additional decorative detail.
Examples:
Instead of:
> An abandoned house.
Describe:
> An abandoned house with broken shutters, vines climbing the walls, dust coating the furniture, and collapsed sections of the roof.
The scene tells a story.
---
# Rule — Let the Environment Support the Narrative
Rather than stating every piece of backstory directly, allow the environment to communicate it visually.
Examples:
Ancient civilization:
Weathered statues.
Broken roads.
Overgrown ruins.
Fresh battlefield:
Scorched earth.
Discarded weapons.
Lingering smoke.
Fishing village:
Wet docks.
Drying nets.
Weathered wood.
The setting should reinforce the intended interpretation.
---
# Observed Behavior — Physical Logic Increases Believability
Seedream and Seedance consistently appear to produce more convincing results when prompts describe physically connected scenes rather than isolated visual elements.
This does not require exhaustive detail.
Often, a few carefully chosen consequences are enough to unify the composition.
---
# Part 3 Summary
This chapter introduced the physical reasoning principles of Prompt Bible V3:
* Every action creates consequences.
* Build motion using State → Action → Consequence.
* Motion propagates into surrounding objects.
* Materials should behave according to their properties.
* The environment participates in the scene.
* Commit to a consistent camera viewpoint.
* Let environments communicate narrative visually.
Together, these principles transform static descriptions into scenes that feel connected, dynamic, and physically coherent.
---
## Author's Note
This chapter marks an evolution from descriptive prompting to simulation prompting.
Rather than asking, "What should be visible?", V3 asks, "What would naturally happen?"
That shift often produces scenes that feel less like staged illustrations and more like captured moments.
---
# Prompt Bible V3 Core
## Part 4 – Seedream / Seedance Execution Layer
### Model-Specific Observations & Practical Rules
---
# Introduction
Parts 1–3 established the core methodology:
* Preserve intent.
* Build clear identities.
* Describe relationships.
* Respect physical logic.
These principles apply broadly.
This chapter focuses specifically on Seedream and Seedance behavior observed during testing.
These are not immutable laws.
Models change.
Training changes.
Future versions may behave differently.
Therefore, this section distinguishes between:
* Core Principles — foundational prompting concepts.
* Observed Behaviors — repeated patterns seen in Seedream/Seedance.
* Practical Rules — methods for working with those behaviors.
The purpose is not to describe what the model "is."
The purpose is to document how to achieve more reliable results with current Seedream/Seedance behavior.
---
# Observed Behavior — Identity Drift
One of the most common challenges in image and video generation is identity drift.
A model may correctly interpret a character initially, then gradually lose defining traits.
This becomes especially noticeable with:
* multiple images
* multi-shot video
* recurring characters
* complex scenes
The model is constantly rebuilding its interpretation.
If identity anchors become weaker than competing information, the subject may shift.
---
# Rule — Reinforce Identity Anchors
For important characters, repeat the information that defines them.
Prioritize:
* silhouette
* unique physical traits
* signature clothing
* distinctive colors
* important accessories
* recognizable behavior
Do not rely on a name alone.
A recurring character should remain recognizable even if the name is removed.
---
# Video Continuity Rule
For multi-shot Seedance prompts:
Explicitly establish continuity.
Example:
> All shots depict the same individual character with identical appearance, clothing, and distinctive features.
This reduces the chance that each shot is interpreted as a separate generation problem.
---
# Observed Behavior — Similar Subjects Increase Confusion
When multiple characters share:
* similar appearance
* similar clothing
* similar descriptions
* similar names
the model may merge identities or assign traits incorrectly.
This becomes more likely as scene complexity increases.
---
# Rule — Give Characters Distinct Anchors
Different characters should have clearly separated identities.
Useful separators:
* different silhouettes
* different color palettes
* different roles
* different behaviors
* different relationships
Example:
Weak:
> Two warriors fighting.
Stronger:
> A tall armored knight in silver plate fighting a shorter rogue wearing dark leather armor.
The model receives clearer assignments.
---
# Observed Behavior — Emotional Framing Has Strong Influence
Seedream and Seedance appear highly responsive to emotional framing.
However, emotion words alone are often weaker than visible emotional evidence.
A prompt saying:
> angry character
leaves interpretation open.
The model must decide:
* facial expression
* posture
* body language
* situation
A stronger approach is to define the visible result.
---
# Rule — Build Emotion Through Evidence
Instead of:
> A terrified person.
Use:
> A person with widened eyes, tense shoulders, defensive posture, and a fearful expression while backing away from the approaching creature.
The emotion is no longer abstract.
It becomes a physical state.
---
# Observed Behavior — Framing Affects Interpretation
The same concept can produce different results depending on how it is framed.
Compare:
> A person trapped in a dangerous situation.
versus:
> A person bravely navigating a dangerous environment.
The physical scenario may be similar.
The interpretation is different.
Seedream/Seedance appear sensitive to the narrative framing surrounding events.
---
# Rule — Prefer Intended Experience Over Unwanted Emphasis
When a scene involves danger, conflict, fear, or destruction, describe the intended tone.
Ask:
Is this:
* horror?
* adventure?
* action?
* comedy?
* mystery?
* triumph?
The emotional framing should match the desired experience.
---
# Observed Behavior — Attention Competition
Long prompts do not automatically produce better results.
Seedream and Seedance must prioritize information.
When too many equally emphasized concepts compete, lower-priority details may disappear.
Common symptoms:
* missing accessories
* altered clothing
* simplified backgrounds
* ignored actions
* genericized characters
---
# Rule — Protect High-Value Information
Before adding another detail, ask:
"Is this more important than what is already present?"
If not, it may reduce clarity.
Protect:
1. Identity
2. Key relationships
3. Main action
4. Unique visual elements
before adding decoration.
---
# Observed Behavior — Prompt Overload
A prompt can fail not because it lacks information, but because it contains too much competing information.
Warning signs:
* many unrelated style descriptors
* excessive adjectives
* multiple conflicting moods
* several simultaneous actions
* unnecessary background detail
Complexity should serve clarity.
---
# Rule — Compress, Don't Inflate
When a prompt becomes crowded:
Remove repetition.
Combine related ideas.
Replace lists with meaningful relationships.
Example:
Instead of:
> old, ancient, weathered, aged, ruined stone castle
Use:
> an ancient stone castle worn by centuries of weather and decay
The concept remains.
The attention cost decreases.
---
# Observed Behavior — Ambiguity Resolution
Models naturally fill in missing information.
This is useful.
It is also a source of unexpected results.
When something is important, do not leave the model to guess.
---
# Rule — Explicitly Anchor Critical Interpretations
If a detail must survive:
* state it clearly
* connect it to the subject
* reinforce it through visible evidence
Avoid assuming the model shares the user's mental image.
---
# Observed Behavior — Exact Wording vs Intent
Seedream and Seedance do not necessarily require exact phrasing.
A concept can survive through multiple descriptions if the underlying meaning remains consistent.
This is why V3 prioritizes intent over wording.
The model needs the right interpretation, not a specific sentence structure.
---
# Practical Seedream / Seedance Optimization Checklist
Before generation:
## Identity
* Is the subject unmistakable?
* Are defining traits reinforced?
## Relationships
* Are important interactions clear?
* Is scale obvious?
## Motion
* Does the action have consequences?
* Does the environment respond?
## Emotion
* Is the feeling visible?
* Is the framing intentional?
## Complexity
* Are important concepts protected?
* Is anything unnecessary competing for attention?
---
# Part 4 Summary
Seedream and Seedance benefit from prompts that:
* establish strong identity anchors
* maintain continuity
* describe visible emotional evidence
* avoid competing concepts
* prioritize important information
* explicitly resolve critical ambiguities
The central lesson:
The model does not need more information. It needs the right information in the right hierarchy.
---
# Prompt Bible V3 Core
## Part 5 – Workflow, Debugging & Advanced Usage
### From Concept to Reliable Generation
---
# Introduction
The previous chapters established the foundations of V3:
* Part 1: How to think about prompting.
* Part 2: How to preserve identity and meaning.
* Part 3: How to create physically coherent scenes.
* Part 4: How Seedream and Seedance specifically interpret prompts.
This final chapter turns those principles into a repeatable workflow.
The purpose of V3 is not to produce a single good prompt.
The purpose is to create a system for consistently moving from an idea to a result while preserving intent.
---
# Core Principle — Prompting Is an Iterative Translation Process
A generated image or video is not simply the result of a prompt.
It is the result of:
Human idea
↓
Intent extraction
↓
Prompt construction
↓
Model interpretation
↓
Generated output
↓
Evaluation
↓
Refinement
Every failure is information.
A result that misses the target reveals where the translation process broke down.
---
# The V3 Workflow
## Step 1 — Identify the Intent
Before writing a prompt, determine:
* What is the subject?
* What is the most important thing about it?
* What must not be lost?
* What would make the result feel wrong?
This defines the priority hierarchy.
---
## Step 2 — Establish the Core Identity
Define:
* who or what the subject is
* the defining visual traits
* the silhouette
* the important relationships
Do not begin with atmosphere.
A beautiful scene with the wrong subject is still a failure.
---
## Step 3 — Build the Scene
Add:
* action
* environment
* physical relationships
* materials
* camera perspective
The scene should answer:
"What is happening?"
not only:
"What exists?"
---
## Step 4 — Add Style Last
Style is important.
However, style should refine the interpretation rather than replace it.
A strong subject with appropriate style usually works better than a stylish but poorly defined subject.
---
## Step 5 — Perform an Interpretation Check
Before generation, review the prompt from the model's perspective.
Ask:
> "If I only saw this text, what would I expect the model to generate?"
If the answer differs from the intended result, revise.
---
# The Interpretation Integrity Check
A final V3 validation pass.
## Identity
Is the subject unmistakable?
Could the model replace it with a more generic version?
---
## Archetype
Is the defining concept preserved?
Has the prompt accidentally weakened the thing that makes it unique?
---
## Relationships
Are important connections clear?
Does the scene explain how things relate?
---
## Scale
Would the model understand the size difference visually?
---
## Action
Does the action have a believable result?
---
## Complexity
Are important details protected from being drowned out?
---
# Failure Diagnosis
When a generation fails, do not immediately add more detail.
First identify the failure type.
---
# Failure Type: Identity Collapse
## Symptom
The character is technically present but feels generic.
## Common Causes
* identity introduced too late
* insufficient anchors
* archetype overwritten by style
* too many competing traits
## Fix
Strengthen:
* silhouette
* unique features
* defining relationships
---
# Failure Type: Archetype Collapse
## Symptom
The model generates a familiar but incorrect version of the concept.
Example:
A unique fantasy creature becomes a generic dragon.
## Common Causes
* default archetype gravity
* insufficient distinguishing traits
## Fix
Reinforce what separates the intended interpretation from the common one.
---
# Failure Type: Scale Drift
## Symptom
Size relationships are incorrect.
## Common Causes
* relying only on measurements
* lack of environmental references
## Fix
Add:
* objects
* architecture
* perspective
* interactions
---
# Failure Type: Motion Failure
## Symptom
Video feels like disconnected poses.
## Common Causes
* action without consequences
* unclear starting state
* missing environmental response
## Fix
Use:
State → Action → Consequence
---
# Failure Type: Character Merging
## Symptom
Characters exchange traits or become visually similar.
## Common Causes
* overlapping descriptions
* similar colors/clothing
* unclear relationships
## Fix
Increase separation between identities.
---
# Failure Type: Prompt Overload
## Symptom
Many details disappear or become inconsistent.
## Common Causes
* too many competing priorities
* excessive decoration
* redundant wording
## Fix
Compress.
Protect the important concepts.
---
# Refinement Strategy
When improving a prompt, modify one category at a time.
Avoid changing everything simultaneously.
A useful order:
1. Identity
2. Composition
3. Relationships
4. Action
5. Environment
6. Style
This makes failures easier to diagnose.
---
# Advanced Principle — Nuclear Pass and Surgical Pass
These are refinement strategies, not mandatory steps.
---
# Nuclear Pass
Used when the interpretation is fundamentally wrong.
Goal:
Rebuild the prompt around the core intent.
Actions:
* remove unnecessary details
* restate identity
* simplify hierarchy
* restore the main concept
The goal is not refinement.
The goal is recovery.
---
# Surgical Pass
Used when the interpretation is mostly correct.
Goal:
Fix specific problems.
Examples:
* improve clothing accuracy
* correct scale
* strengthen expression
* adjust environment
Only modify the failing component.
---
# Confidence Classification
Not every V3 statement has the same status.
Future versions should maintain this distinction.
---
# Core Principle
Stable methodology.
Examples:
* Intent preservation.
* Hierarchical prompting.
* Relationship anchoring.
* Visual translation.
---
# Observed Behavior
Repeated Seedream/Seedance patterns.
Examples:
* identity drift
* attention competition
* emotional framing effects
* continuity challenges
---
# Experimental Hypothesis
Interesting possibilities requiring further validation.
Examples:
* emerging prompt structures
* possible model-specific optimizations
* unexplained generation patterns
Experiments should inform the system, not silently become rules.
---
# Final V3 Philosophy
Prompt Bible V3 is not about commanding the model.
It is about communicating with it.
The best prompts do not overwhelm the model with instructions.
They provide the right information in the right order so the model can construct the intended interpretation.
The ultimate goal is:
Maximum intent preservation with minimum unnecessary complexity.
A good prompt is not the longest prompt.
It is the prompt that leaves the smallest gap between imagination and generation.
---
# Prompt Bible V3 Core Complete
## Summary of the Five Parts
### Part 1 — Foundations
The philosophy:
* Intent preservation
* Hierarchy
* Attention economy
### Part 2 — Identity & Visual Reasoning
The representation:
* Identity anchors
* Archetypes
* Relationships
* Scale
* Visual evidence
### Part 3 — Motion & Physics
The simulation:
* Consequences
* Materials
* Environment
* Camera
### Part 4 — Seedream/Seedance Layer
The execution:
* Drift prevention
* Continuity
* Emotional framing
* Complexity management
### Part 5 — Workflow & Debugging
The methodology:
* Construction
* Evaluation
* Refinement
* Failure correction