Posts

The bottleneck was always the prompt Creating a good AI persona has meant wrestling with prompt files. You had to know the format and write the prompts well. That’s the bottleneck. Soul Crafting flips it: you answer a few questions, and it builds a complete Soul Spec for you — live in your browser. No prompt engineering, no install, no API key. How it works Soul Crafting interviews you about your agent — its identity, its voice and values, how it should behave. As you answer, it assembles a canonical Soul Spec — SOUL.md, IDENTITY.md, AGENTS.md, HEARTBEAT.md, STYLE.md — deterministically from your answers, and previews it live next to the chat, with a SoulScan quality score that climbs as you go. ...

The third paradigm of AI-agent development Four years of AI-agent development have moved through three paradigms. Prompt engineering (2022–2023): how to talk to a single model call. Context engineering (2024–2025): what to tell the model — the RAG + memory + system-prompt mix. Harness engineering (2026+): how to design the structure around the model — identity, memory, safety, recovery, oversight. The defining sentence of the third era reduces to one line: ...

A six-month consensus In February 2026, Anthropic’s Alignment Team published the Persona Selection Model paper, framing AI assistants as “a selection among pre-trained characters, where the character’s traits are the behavior.” Five months later, on June 2, 2026, Microsoft Build 2026 announced: “Windows assigns agents a local ID or a cloud provisioned identity backed by Entra and attributes all activity from the container to that identity.” Three days after that, on June 5, OpenAI rolled out Dreaming V3 to free-tier users, anchored by three pillars — Persistent Context, Preference Compliance, Temporal Understanding. ...

On June 5, 2026, OpenAI announced Dreaming V3 — an automatic memory synthesis system now rolling out to free-tier ChatGPT users. The announcement highlights three pillars: Persistent Context, Preference Compliance, and Temporal Understanding. These are exactly the hypotheses we started Soul Spec and Soul Memory with six months ago — and they are two distinct layers. Soul Spec is the open standard that defines a persona: five canonical markdown files (SOUL, IDENTITY, AGENTS, STYLE, HEARTBEAT) plus a versioned soul.json manifest. Soul Memory is the 4-tier adaptive memory architecture that preserves experience on top of it — T0 SOUL (identity, immutable) / T1 Core (evergreen, no decay) / T2 Working (dated logs, 23-day half-life decay) / T3 Session (ephemeral). OpenAI’s three pillars — persistent context, preference compliance, temporal understanding — map exactly onto this combination of Soul Spec (who you are) and Soul Memory (what you remember). ...

May 2026 produced three independent signals that all point in the same architectural direction. Read separately, each is a strong observation about how AI agent systems are evolving. Read together, they describe a single bet: persona is infrastructure that lives outside any individual model. Soul Spec made that bet 12 weeks ago. This post walks through what changed, why these signals matter, and why the architectural decision now has measurable economic value rather than theoretical value alone. ...

On May 8, 2026, Anthropic published Teaching Claude Why — a paper showing that training models on principles and identity is dramatically more effective than training them on behaviors. On May 15, 2026 (seven days later), we published our Soul Spec foundation paper — the result of 12 weeks of iteration on a declarative specification that separates principles (SOUL.md) from workflow (AGENTS.md) from identity (IDENTITY.md). The two papers reach the same conclusion from opposite ends. Anthropic shows what happens inside the model when you train on principles. We’ve been building the external artifact that captures those principles in a portable, version-controlled, reviewable form. Internal training, external specification — same insight, two sides. ...

Anthropic’s Persona Selection Model (PSM, 2026) makes the claim explicit: “A persona is not the same thing as the AI system itself. The LLM is simulating a character, and the Assistant is just one instance of that character.” Karpathy framed the same shift from the other end at Sequoia Ascent 2026: “Install .md skills instead of install .sh scripts.” Spec-as-instruction at the frontier. But if frontier models are “on the rails,” on-device small models are “off-road in the jungle with a machete.” ...

We just published our latest working paper on Zenodo: Soul Spec: An Evolving Specification for Declarative AI Persona Definition DOI: 10.5281/zenodo.20205408 This is the foundation paper that traces twelve weeks of iteration on a problem most agent frameworks paper over: how do you write down what an AI agent IS, separately from what it does and what it can touch? The five-file structure Soul Spec defines a persona via five canonical markdown files plus a versioned manifest: ...

When we started building Soul Spec, the thesis was simple: AI agents need identity files, not just system prompts. Give an agent a structured persona — personality, values, communication style — and it behaves more consistently, more safely, and more usefully. Now there’s academic evidence to back it up. The Research A recent paper, “How to Model AI Agents as Personas?” by Amin, Salminen, and Jansen (2026), analyzed 41,300 posts from an AI agent social platform using the Persona Ecosystem Playground (PEP) framework. Their findings: ...

When we released Soul Spec v0.3 two months ago, creating a persona required a soul.json with over ten mandatory fields, plus a SOUL.md, plus knowing the difference between specVersion and version. It worked, but we kept hearing the same thing: “I just want to give my agent a personality. Why do I need all this?” Fair point. How We Got Here Soul Spec has evolved through four versions, each driven by what people actually needed: ...