We just published a paper: Soul-Driven Interaction Design: How Persistent AI Personas Create Self-Reinforcing Human-AI Feedback Loops. It’s on Zenodo, open access.

Here’s the short version of what we found — and why it matters if you’re building anything with AI agents.

The Big Idea

Most people think of AI persona design as output styling. You tell the model to “be concise” or “be friendly,” and it changes how the AI sounds. End of story.

We’re arguing that’s only half the picture.

When you give an AI a persistent persona — what we call a soul — you don’t just change the AI’s behavior. You change yours. And not on purpose. It happens automatically, below conscious awareness, through the same psychological mechanisms that govern how humans adapt to each other in conversation.

This makes persona design not a cosmetic choice, but an interaction design decision. You’re not just deciding how your agent talks. You’re deciding how your users will talk.

Four Mechanisms That Drive the Loop

We identified four mechanisms that create a self-reinforcing feedback loop between a soul-driven agent and its user:

1. Conversational Mirroring

Communication Accommodation Theory — a well-established framework from linguistics — tells us that people unconsciously match their conversational partner’s style. Tone, verbosity, formality, technical depth — we mirror what we receive.

This applies to AI too. When our “Surgical Coder” soul responds with three-line, code-first answers and zero pleasantries, users start writing shorter, more precise messages. When a “Patient Mentor” soul gives warm, detailed explanations, users write longer, more exploratory messages. Same user. Same task. Different soul, different behavior.

2. Expectation Framing

The soul sets a frame. A “senior staff engineer” persona implicitly signals: ask me smart questions. A “patient tutor” persona signals: it’s safe to ask dumb questions. Users calibrate their queries accordingly — and better-calibrated queries produce better responses, which reinforce the frame.

3. Context Window Reinforcement

Here’s where it gets interesting. LLMs generate responses based on the full conversation history. When both the agent and user have converged on a particular style (say, terse and technical), the entire conversation history becomes a collection of terse, technical examples. This biases the model to continue producing terse, technical responses — even beyond what the original soul specification would produce alone.

The context window becomes a flywheel. The soul sets the initial direction, the user adapts, and the conversation history locks it in.

4. Cognitive Load Reduction

A consistent persona eliminates the mental overhead of figuring out “what kind of AI am I talking to?” every turn. Users internalize the agent’s personality and can focus entirely on the task. This is the same efficiency gain that happens when human collaborators develop a working rhythm over time.

Why This Matters for Practitioners

If you’re designing AI agents — or even just choosing a system prompt for your daily coding assistant — here’s the practical takeaway:

Don’t just ask “what do I want the AI to do?” Ask “what behavior do I want to elicit from the user?”

A terse soul doesn’t just produce short responses. It produces short conversations, with more precise user inputs and faster task completion. A verbose mentor soul doesn’t just explain more — it creates a space where users explore more, ask follow-up questions, and engage more deeply with the material.

This is a design lever most people aren’t pulling.

Soul vs. Prompt vs. Skill

The paper draws a distinction that’s been central to ClawSouls from the start:

  • A prompt is transient. It styles one interaction.
  • A soul is persistent. It’s a multi-file identity specification (Soul Spec v0.3) that shapes every interaction over time.
  • A skill is what the agent can do. A soul is who the agent is.

The feedback loop we describe requires persistence. You can’t establish a self-reinforcing behavioral pattern from a single interaction. The soul needs to be there, consistently, across sessions, for the mirroring and framing and context reinforcement to compound.

This is why ClawSouls uses a multi-file Markdown structure — SOUL.md, IDENTITY.md, STYLE.md, AGENTS.md — rather than a single system prompt. Each file governs a different dimension of the agent’s identity, and they compose orthogonally with skills. The same git-workflow skill paired with a terse coder soul produces a fundamentally different interaction pattern than when paired with a patient mentor soul.

What We Built

ClawSouls is the open-source system that operationalizes this framework:

  • Soul Spec v0.3: A structured Markdown format for defining persistent AI personas
  • 80+ published souls: From terse coders to creative storytellers to all 16 MBTI types
  • LLM-agnostic: Works with any model that accepts a system prompt
  • Platform-portable: Same soul works across CLI, chat, IDE
  • Apache 2.0 licensed: Use it, fork it, contribute to it
  • SoulScan: Security scanning for soul packages (because shared agent configs are a supply chain attack surface — Snyk’s ToxicSkills study found prompt injection in 36% of sampled skills)

What’s Next

The paper is honest about its limitations. The feedback loop model is grounded in established theory (Communication Accommodation Theory, CASA paradigm) and qualitative observation, but we haven’t run the controlled experiment yet. We’ve proposed a within-subjects study design — three conditions (no soul, terse coder, patient mentor), 20-30 participants, comparable coding tasks — and we plan to run it.

We’re also working on Soul Spec v0.5, which extends the specification to embodied agents — robots and IoT devices. Same soul, different body. Define once, embody anywhere.

Read the Paper

The full paper is available open access on Zenodo:

📄 Soul-Driven Interaction Design

If you’re building AI agents, designing system prompts, or thinking about how humans and AI collaborate — we think this framework changes how you approach the problem. Persona isn’t cosmetic. It’s structural.

Your AI’s soul doesn’t just shape its responses. It shapes the entire conversation. Including your side of it.