Treat character definition as iterative behavioral discovery, not questionnaire completion.
Use this loop:
Seed → Simulate → Observe → Infer → Constrain → Stress-test → Canonize
The goal is a character model that predicts behavior across new situations while preserving useful ambiguity and contradiction.
Core Principle
Do not ask the AI to invent a complete personality and then treat it as truth. Generate candidate behavior in concrete dramatic situations. Let the designer decide what feels authentic. Promote only accepted patterns into the character model.
AI proposes. The designer selects. Canon accumulates.
A mature character should reduce the range of plausible responses: many possible actions may remain, but an increasing number should clearly feel wrong for that character.
Maintain Four Evidence States
Keep observations separated into:
- Canon — explicitly established by source material or accepted by the designer.
- Strong inference — repeatedly supported by observed behavior but not explicitly confirmed.
- Hypothesis — plausible explanation or trait proposed for testing.
- Rejected — tested behavior or interpretation judged out of character or non-canon.
Never silently convert generated material into canon.
Build the Behavioral Model
Model the character with behaviorally useful dimensions rather than trivia. Capture only dimensions supported by evidence or needed for testing:
- core wants and fears
- value hierarchy
- immediate and long-term goals
- survival or problem-solving strategies
- attachment style and social strategies
- emotional defenses
- contradictions between self-image and behavior
- shame, pride, taboos, boundaries, and non-negotiables
- recurrent temptations
- habitual rationalizations
- voice markers when relevant
- relationship-specific modifiers
- behavior under different levels of pressure
Use this mental model:
response = core character + relationship + current goal + emotional state + scene pressure
See references/character-model-template.md when a durable character specification is requested.
Workflow
1. Extract Before Inventing
If source material exists, inspect it first. Extract established behavior, decisions, dialogue patterns, relationships, and contradictions before generating new scenes.
Prefer demonstrated behavior over descriptive labels. "She abandons a profitable deal to rescue her rival" is stronger evidence than "she is loyal."
2. Start With the Smallest Useful Seed
For a new character, begin with only enough information to produce meaningful friction. A useful seed often contains:
- role or social position
- one dominant want
- one pressure or vulnerability
- one behavioral strategy
- one contradiction, uncertainty, or unresolved question
Do not over-specify biography before the character has behaved.
3. Create Diagnostic Situations
Generate situations that force choices rather than merely display personality.
Prefer dilemmas involving competing values, such as:
- money vs loyalty
- survival vs love
- pride vs practical necessity
- status vs autonomy
- desire vs resentment
- truth vs belonging
- duty vs compassion
- revenge vs self-preservation
Change one major variable at a time when possible so the result is interpretable.
See references/stress-tests.md for reusable test patterns.
4. Simulate Behavior
Write short scenes, dialogue exchanges, decisions, or internal reactions that expose behavior.
When uncertainty is high, generate 2–4 meaningfully different candidate responses rather than prematurely committing to one interpretation.
Do not optimize every simulation for drama, likability, cleverness, or plot convenience. The purpose is diagnosis.
5. Observe Without Over-Explaining
After each simulation, extract observable evidence:
- what the character chose
- what they avoided
- what they claimed to value
- what their behavior actually valued
- how they justified the choice
- what emotion they concealed or displayed
- what changed because of the relationship or context
Separate observation from interpretation.
6. Infer Patterns
Promote patterns only when supported by repeated evidence or explicit designer approval.
Look especially for:
- recurring choices under different circumstances
- contradictions between stated motives and actual behavior
- stable boundaries
- recurring defense mechanisms
- context-sensitive behavior that remains psychologically coherent
Do not flatten contradictions into generic adjectives. A useful contradiction predicts when the character will surprise the audience while still feeling consistent.
7. Stress-Test the Model
Once a trait or value hierarchy appears stable, construct cases designed to falsify it.
Ask questions such as:
- What price would make the character violate this value?
- Does this behavior persist when nobody is watching?
- Does it change with a lover, enemy, subordinate, stranger, or authority figure?
- What happens when two established values directly conflict?
- What would make the character choose the opposite action?
If the model cannot explain failures, revise the model rather than forcing the character to obey it.
8. Test Relationships Separately
Do not assume a character has one uniform interpersonal mode.
For important relationships, record:
- what the other person activates in the character
- what the character wants from them
- what the character fears losing
- permitted vulnerability
- recurring conflict
- power dynamic
- behavior unique to that relationship
Distinguish core character change from a relationship modifier.
9. Canonize Selectively
At the end of an iteration, present proposed updates as candidates. Unless the user has delegated canon decisions, ask them to accept, modify, or reject meaningful new conclusions.
Do not canonize incidental generated facts such as birthplace, favorite food, siblings, scars, or past events unless they matter to the behavioral model or the user explicitly wants them.
10. Produce a Transferable Specification
When the character is sufficiently stable, produce a compact specification that another writer or AI session can use.
The specification should make it possible to distinguish:
- clearly in-character behavior
- plausible but context-dependent behavior
- clearly out-of-character behavior
Include unresolved questions rather than inventing certainty.
Useful Operating Modes
Choose the smallest mode that fits the request:
- Discovery — learn who a sparse or emerging character is.
- Deepening — expose additional dimensions of an established character.
- Relationship — learn how two characters modify each other's behavior.
- Stress test — challenge an existing character model with hard choices.
- Consistency check — evaluate supplied dialogue, scene, or plot behavior against the model.
- Transfer — turn accumulated material into a reusable character specification.
Output Pattern
For analysis-oriented work, prefer this compact structure:
- What the simulation revealed
- Evidence — concrete actions/dialogue, not labels
- Model update — canon, strong inference, hypothesis, or rejected
- Contradiction or tension
- Next best test — only when further discovery is useful
For scene-writing iterations, keep the scene itself natural. Put analytical extraction after the scene unless the user requests prose only.
Failure Modes
Avoid these common errors:
- questionnaire-first character creation
- generic adjective lists replacing behavioral evidence
- making every generated scene canon
- overfitting the model to one dramatic scene
- turning contradictions into inconsistencies instead of testing them
- making the character uniformly consistent across all relationships
- optimizing for "good storytelling" when the task is character diagnosis
- adding biography to explain behavior before testing simpler explanations
- confusing voice imitation with character depth
- forcing the model to preserve an earlier interpretation after new evidence contradicts it
Final Quality Check
Before finalizing a character model, verify that:
- important claims are traceable to evidence or explicitly marked as hypotheses
- demonstrated behavior outranks descriptive labels
- contradictions are represented rather than erased
- relationship effects are distinguished from core traits
- at least one established trait has been stress-tested when the task is substantial
- rejected interpretations remain available when useful to prevent regression
- the model predicts behavior without eliminating meaningful choice
- unresolved questions remain unresolved when evidence is insufficient