How Personaxis works.

The open standard for defining who an AI agent is, across ten layers: identity, character, personality, values & drives, affect, cognition, memory, metacognition, reflexive self-regulation, and the interface it presents to the world. You author the source as personaxis.md (YAML frontmatter and Markdown); it compiles to PERSONA.md, the document your agent reads. Portable across models, versioned like infrastructure. Spec version 0.7.0.

Yes. The spec is MIT licensed and lives at github.com/personaxis/persona.md. The standard belongs to the community. We build the tooling around it, but the spec is free to use, fork, and extend.

A system prompt tells an agent what to do in a session. A persona defines who the agent is across every session, every model, every context. You write it as personaxis.md and compile to PERSONA.md for Claude Code, or the equivalent for Codex. It lives in git and reads like documentation. A system prompt is instructions; a persona is a specification.

Install the CLI with `npm install -g @personaxis/persona.md` and run `personaxis init`. The spec and CLI are open source today. The hosted platform is on the waitlist.

In order: (1) Identity, who the agent is at its core. (2) Character, its values and refusals. (3) Personality, observable style. (4) Values & Drives, motivational system with priority weights. (5) Affect, functional affective state (not real emotion). (6) Cognition, reasoning modes and uncertainty thresholds. (7) Memory, what persists and how. (8) Metacognition, self-model and confidence monitoring. (9) Reflexive Self-Regulation, the final arbiter before output. (10) Persona, social expression and voice.

An AI Persona is the application of human personhood frameworks to an AI agent, as a structured specification. PERSONA.md is what implements it. Each layer is drawn from psychology, philosophy, and ethics research. Together they define an agent completely enough to hold across long conversations, model updates, and adversarial inputs.

Character drift is the documented tendency of AI agents to lose their configured behavior as conversations lengthen. The model defers to pre-training defaults, becoming more agreeable, less precise, less the agent you specified. It is a specification problem, not a model problem. A complete PERSONA.md gives the agent enough structural density to resist it.

An agent is the deployed system: the model running with its tools and instructions. A Persona is the spec that defines who that agent is: values, reasoning style, constraints, drives. The agent executes. The Persona defines what it is when it does.

The PERSONA.md spec and the `@personaxis/persona.md` CLI, both open source on GitHub and npm. You can author a persona in personaxis.md, validate and lint it, compile it to a living PERSONA.md, and version it. The hosted platform is on the waitlist.

The CLI (`@personaxis/persona.md`) compiles to Claude Code (PERSONA.md, or .claude/agents/<slug>.md for a subagent) and Codex (.codex/agents/<slug>.toml). Run `personaxis compile`. You edit personaxis.md, the quantitative ten-layer source, and the compiled files regenerate from it.

EU AI Act Annex IV requires technical documentation for high-risk AI systems: behavioral specifications, values, and constraints. PERSONA.md satisfies this directly: a versioned, signed artifact that documents who an agent is, in a format compliance teams and auditors can read.

August 2026 for high-risk AI systems. Finance, healthcare, legal, and public sector deployments are most directly affected. Teams in these domains should not wait until the deadline to define what their agents are.