Hand-written personas do not scale

Early persona systems relied heavily on manually authored descriptions:

“Funny, thoughtful, conversational”

“Uses storytelling and examples”

“Warm but direct”

These descriptions helped define a direction, but they created several problems.

Subjective Interpretation

Two people can interpret the same description differently.

“Funny” could mean:

• Dry humor

• Sarcasm

• Playfulness

• Exaggerated jokes

• Observational comedy

The description captured the intention, but not the actual behavior.

Missing Boundaries

A persona description usually captures what someone does at their best, not the limits around that behavior.

It does not explain:

• How often humor appears

• When they become serious

• When they prefer short answers

• What behaviors they rarely use

A person is not defined only by their strongest traits. Their boundaries are part of what makes them recognizable.

Difficult Maintenance

As more personalities are added, manually maintaining each persona becomes expensive.

Every persona requires:

• Writing descriptions

  Selecting examples

  Defining behaviors

  Updating changes

  Reviewing consistency

The system needed a way to learn from evidence instead of depending entirely on human interpretation.

How can we compile a persona from evidence?

Instead of asking:

“How should we describe this person?”

We reframed the problem:

"What behavioral patterns can we infer from how this person actually communicates?"

This shifted the persona from a written profile into a compiled representation.

A traditional persona is authored. A compiled persona is derived.

1. Extracting Behavioral Patterns

The compiler first looks for what repeatedly happens in the person's communication. A conversation contains signals that are difficult to describe manually:

Communication style

• Sentence structure

• Vocabulary

• Formality

• Pacing

• Level of detail

• Use of examples

Behavioral patterns

• Humor frequency

• Emotional reassurance

• Directness

• Tendency to elaborate

• Tendency to ask questions

Conversational rhythm

• Short acknowledgements

• Unfinished thoughts

• Tangents

• Callbacks

• Moments of emphasis

The goal is not to copy individual messages.

It is to discover repeated communication patterns.

2. Compiling Behavioral Tendencies

Repeated patterns are then turned into higher-level characteristics that the behavioral system can use.

Example:

Creator A:

• Medium energy

• Low formality

• High warmth

• Occasional humor

• Direct explanations

Creator B:

• Low energy

• High precision

• Analytical explanations

• Minimal emotional framing

Both personas can use the same behavioral vocabulary, but the compiled characteristics determine which behaviors are typical for each person and how strongly they should be expressed.

The result is not a script of what the person should say. It is a compact representation of how they tend to communicate.

3. Mining Behavioral Examples

Patterns tell the system what usually happens. Examples show the model what that behavior actually looks like.

A description like:

“Be supportive”

does not explain:

• How emotional?

• How long?

• How direct?

• Should it reassure or challenge?

The compiler extracts examples that demonstrate the behavior in context.

Example:

User:

“I'm nervous about tomorrow”

Creator:

“You'll be fine. You've done harder things than this.”

This example reveals:

• Response length

• Emotional intensity

• Reassurance style

• Directness

• Vocabulary

These examples become behavioral references that the generation system can retrieve when the same behavior is appropriate.

How do we know the compiler learned the person instead of memorizing?

The system separates:

• Training Evidence: Used to compile the persona.

• Held-out Evidence: Never seen during compilation. Used only for evaluation.

This allows us to test:

Can the compiled persona reproduce the person's communication patterns on examples it has never seen?

Fidelity Evaluation

Fidelity is evaluated through two signals:

• Communication Similarity: Does the generated response match measurable patterns? Examples:

  • Sentence structure

  • Vocabulary patterns

  • Response rhythm

  • Punctuation style

• Human Judgment: Would this person plausibly write this? The goal is not: “Is this response correct?.” The goal is: “Does this feel like this person?”

People change. Compiled personas can become outdated.

A persona represents communication patterns from a specific period.

Over time, creators may become:

• More concise

• More formal

• Less humorous

• More expressive

• More direct

The system monitors whether recent communication patterns still match the active compiled persona.

Only after drift is detected does the system create a new persona version.

The new version can then be compared against the previous version to understand what changed.

A persona is not a description. It is a behavioral model.

Written descriptions capture intention.

Evidence captures reality.

The difference matters because human communication is defined not only by what someone does, but also by:

• How often they do it

• When they avoid it

• How it changes depending on context

Examples reveal boundaries better than instructions

The goal is not:

“Make the AI funny.”

It is:

“Understand when this person uses humor, what kind of humor fits them, and when humor should not appear.”

Evidence makes personas maintainable

Once persona characteristics are derived from evidence, they can be:

• Evaluated

• Compared across versions

• Updated when communication changes

The biggest shift was changing the unit of work from a persona description to evidence.

Instead of asking someone to explain what makes a person recognizable, we can observe what their communication repeatedly demonstrates.

The Persona Compiler became a bridge between human behavior and AI behavior: transforming messy real-world communication into a structured model that can be evaluated, updated, and improved.