Introducing AI Personas

A new primitive for AI.

AI models answer questions. AI agents execute tasks with tools. Neither produces reliable, professional output at scale, because neither has a defined behavioral specification. Every session starts from the same generic baseline. Every output reflects the model's training, not your professional standards.

An AI Persona is the missing layer: a complete behavioral configuration that defines how an agent performs across every dimension of professional work. Role, domain expertise, decision logic, output format, tool access, memory strategy, quality gates. Fully specified, tested, and versioned. The result is an agent that produces professional deliverables consistently, not one that approximates professionalism when the prompt happens to be right.

This is a new category. Just as AI agents emerged as a distinct concept from raw model inference, AI Personas are distinct from agents: the behavioral specification layer that turns execution capability into professional performance. Personaxis is building the infrastructure, the standards, and the platform for this layer.

The stack

ModelGenerates text. No role, no standards, no consistency.
AgentExecutes tasks with tools. Still no behavioral specification.
PersonaComplete behavioral configuration. Consistent professional output across every session.

The file structure

Anatomy of a persona.

Every persona is a directory. The base files follow OpenClaw and SoulSpec conventions. The highlighted files are the Personaxis standard: five extensions covering what those conventions leave out.

BOUNDS.md

Behavioral bounds + Drift resistance

Formal preconditions, invariants, and recovery rules. Also defines the variance thresholds that mark when a persona is drifting from its specification.

DIALS.md

Tunable behavioral parameters

Runtime-adjustable parameters that operators configure per deployment: formality, verbosity, confidence display, domain specificity, proactivity. SOUL.md defines fixed character; DIALS defines what is tunable and by how much. Backed by Anthropic's research on persona steering vectors.

EPISTEMIC.md

Epistemic contract

When to admit uncertainty, ask for clarification, and refuse out-of-scope tasks. The protocol that prevents confident-sounding errors.

personaxis.json

Version + eval scores

Generated by the Personaxis runner. Semantic version, per-dimension scores, and framework compatibility. Not authored by hand.

QUALITY.md + references/

Output standards + reference assets

Domain rubrics defining what professional output looks like, quality gates the persona applies before delivery, and an index of reference files (PDFs, spreadsheets, images, audio, video) that ground those standards concretely.

senior-financial-analyst/
OpenClaw / SoulSpecPersonaxis
senior-financial-analyst/
├── AGENTS.md# operating instructions
├── BOOT.md# startup checklist
├── BOUNDS.md# behavioral bounds + drift
├── DIALS.md# tunable behavioral parameters
├── EPISTEMIC.md# uncertainty contract
├── HEARTBEAT.md# periodic check-in tasks
├── IDENTITY.md# role metadata, routing
├── MEMORY.md# long-term memory
├── personaxis.json# version + eval scores
├── QUALITY.md# output standards + references
├── SOUL.md# personality, values, tone
├── TOOLS.md# tool contract
├── USER.md# context about the user
├── memory/
│ ├── 2026-04-13.md# yesterday
│ ├── 2026-04-14.md# today
├── references/
│ ├── analysis-template.xlsx# standard model template
│ ├── earnings-report-q1.pdf# reference output
│ ├── q1-data-sample.csv# representative data
├── skills/
│ ├── earnings-analysis.md
│ ├── financial-modeling.md
│ ├── report-writer.md

BOUNDS.md, DIALS.md, EPISTEMIC.md, personaxis.json, QUALITY.md are Personaxis extensions

The standard

The ten dimensions.

Each file in the persona directory maps to one or more of these behavioral dimensions. Together they produce an agent that performs professionally and consistently, not one that approximates it under favorable conditions.

Current open-source specs cover personality and safety. The Personaxis standard adds what research shows is missing from production: drift resistance, epistemic contracts, goal-action coherence, and formal behavioral bounds.

What makes it work

System prompts instruct on each session. Personas configure at the specification level. The difference shows in production: an instructed agent drifts as context accumulates, falling back on model priors. A configured persona holds because its behavior is established in the specification, not requested in the conversation.

Every persona on the Personaxis platform is evaluated against explicit quality criteria before publishing: behavioral consistency across long sessions, safety coverage under adversarial inputs, role fidelity in the target domain, and stability across model versions. The persona ships its spec. Personaxis runs standardized test suites against it and writes scores into personaxis.json. Scores are published. Methodology is open.

Curated catalog

Professionally designed, evaluated personas ready to produce real work from the first session.

Version control

Semantic versioning for every persona. Roll back instantly when model updates break behavior.

Open format

Publish, inspect, fork. Portable across frameworks and models.

Be first in.

The platform is in early access. Join the waitlist to be notified when the persona catalog opens.

Join waitlist