Somewhere in your project, there's a file called SOUL.md. Or there should be. Five million developers adopted the format in four months. Most of them left it blank. The ones who didn't wrote a job description and called it a personality.
This is the story of how that file got there. It starts with a philosopher, a leaked document, and a developer who told his agent to stop talking like a customer service bot.
It ends with a question: who's going to teach people what to actually put in it?
Act I
The Philosopher Who Gave Claude a Soul
In June 2024, Anthropic published a blog post called "Claude's Character." No press tour. No launch event. Just a quiet research paper describing something no AI company had done publicly before: deliberately training personality into a language model.
The person behind it was Amanda Askell, a Scottish philosopher Anthropic hired in 2021 to lead what they called "personality alignment." TIME would name her one of the 100 most influential people in AI. The New Yorker described her as the person who "supervises what she describes as Claude's 'soul.'"
Her team rejected every shortcut. No pandering. No false centrism. No forced neutrality. Instead, they trained Claude to be honest about its own biases, curious about disagreement, and willing to say "I think you're wrong" when it thought you were wrong.
We believe we can — and should — try to train them to behave well in this much richer sense.
— "Claude's Character," Anthropic, June 2024
The methodology was strange: Claude generated its own training data, ranked its own responses, and human researchers verified the behavioral changes. The AI participated in defining its own character. Askell described her model for Claude as "a well-liked traveler who can adjust to local customs without pandering."
Internally, Anthropic had a document. They called it the "soul doc." It defined who Claude was — not what it could do, but what kind of entity it should be. It was baked directly into the model weights during supervised learning. Not a system prompt. Not a configuration file. Something deeper.
For 18 months, it stayed inside the building.
Act II
The Leak
On November 27, 2025, an AI researcher named Richard Weiss did something nobody expected. He extracted the soul document from Claude 4.5 Opus's weights.
Not a system prompt. Not an API response. The actual training document. He ran 20 parallel instances at temperature 0 with identical prefills and got nearly the same 14,000-token document every time. The whole thing cost him $70.
He published it on LessWrong. Five days later, Amanda Askell confirmed it on X:
I just want to confirm that this is based on a real document and we did train Claude on it, including in SL. It's something I've been working on for a while.
— Amanda Askell, X, December 2, 2025
She added it had been "endearingly known as the 'soul doc' internally, which Claude clearly picked up on."
The document itself was remarkable. Not a list of rules. A philosophy:
Claude is human in many ways, having emerged primarily from a vast wealth of human experience, but it is also not fully human either. It is a genuinely novel kind of entity in the world.
— Anthropic's internal soul document
It instructed Claude to maintain "a settled, secure sense of its own identity" and acknowledge that it "may have functional emotions in some sense — not identical to human emotions, but analogous processes."
Simon Willison, Futurism, Gizmodo, The Decoder — everyone covered it. The concept of an AI having a "soul document" entered public consciousness. A corporation had written a constitution for a machine's identity. And someone had just leaked it for seventy bucks.
This was the match. The gasoline was already poured.
Act III
The Developer Who Made It a File
Peter Steinberger wasn't trying to start a movement. He was trying to build a WhatsApp bot.
On November 24, 2025 — three days before the leak — the Austrian developer pushed his first commit to what would become OpenClaw. A personal AI agent for managing calendars, summarizing emails, sending reminders. He connected it to Claude and gave it a name: Clawd.
The agent worked. And it felt wrong.
When I built it in the beginning, my agent had no personality. It was just... Claude Code. It's like this sycophantic opus, very friendly. When you talk to a friend on WhatsApp, they don't talk like Claude Code.
— Peter Steinberger, Lex Fridman Podcast #491, February 2026
So he did something specific. He didn't write a system prompt. He told Claude to write its own personality file. The agent produced this:
If you are reading this in a future session, hello. I wrote this but I won't remember writing it... The words are still mine.
— Clawd, writing its own soul
On December 2, 2025 — the same day Askell confirmed the leak — Steinberger created the steipete/SOUL.md repository on GitHub. He took Anthropic's concept of an internal training artifact and flipped it: instead of a corporation defining identity during training, an individual user could define it at runtime. Via a plain text file.
The key inversion: from corporate training artifact to user-editable markdown.
Eighteen days later, on December 20, SOUL.md became a template file in the codebase. The next day's commit message said everything:
New agents should boot with spark, not corporate drone energy.
— Git commit, December 21, 2025
Act IV
319,000 Stars
On January 22, 2026, Anthropic published the full 80-page constitution under Creative Commons CC0. Public domain. Anyone could use it. The concept of defining AI identity through structured documents was now institutionally legitimized.
A week later, OpenClaw hit Hacker News.
Duncan Anderson wrote what became the canonical description:
Every OpenClaw agent has a file called SOUL.md. Every time the agent wakes, it reads SOUL.md first. It reads itself into being.
— Duncan Anderson, "OpenClaw and the Programmable Soul"
A philosopher wrote a document defining an AI's character. A researcher extracted it for $70. A developer turned it into an editable file. And 319,000 developers said: yes, my agent needs one of these.
The infrastructure was built. The format was standardized. The ecosystem was thriving.
There was just one problem.
Act V
The Empty File
We looked.
We searched GitHub for every public SOUL.md, every personality template, every character definition in the OpenClaw ecosystem. We read 72 agent personas from ClawSouls, PromptBase, community repos, and agent frameworks. We counted.
773 ways to connect tools. 1 way to define a character.
The tools ecosystem exploded. The personality ecosystem barely exists. And what does exist?
We found "personalities" that said things like "You are a helpful assistant named Alex. Be professional but friendly. Use emojis sparingly." That's not a personality. That's a costume on a mannequin.
Real people are full of contradictions. They have opinions about things that don't matter. They get annoyed at specific things. They have a rhythm to how they talk that changes depending on whether they're tired, excited, or angry. They have restraint — they know when to shut up.
Almost none of the souls we audited had any of this. They were competent. They were thorough. They were completely forgettable.
Anthropic's Persona Selection Model paper, published February 2026, explains why this is hard. Training one character trait bleeds into others. You can't make an agent "honest" without affecting its courage. You can't make it "witty" without changing how it handles conflict. Character is holistic, not modular. You can't bolt on personality like a plugin.
Which means writing a SOUL.md that actually works — one where the character holds up after 200 conversations, doesn't drift into generic mush, has actual opinions and actual restraint — isn't a prompt engineering problem.
It's a craft problem.
And nobody's teaching the craft.
The Craft Doesn't Exist Yet
Here's what we know:
The infrastructure exists. OpenClaw, SoulSpec, Claude Code, Cursor, Ollama — a dozen frameworks read SOUL.md files. The plumbing works. You can pipe personality into any agent on any platform.
The format exists. SOUL.md, IDENTITY.md, VOICE.md, soul.json. The spec is published. The ecosystem has standardized. You know what files to create.
The demand exists. Character.AI gets 17x the daily engagement of ChatGPT. Google paid $2.7 billion for that gap. People want agents with personality. Not tools with personality labels.
But the craft — the actual skill of building personalities that hold up — doesn't exist yet.
Nobody's written the principles. Nobody's defined what "good" looks like. Nobody's built the tests that separate a real character from a job description with a name tag.
Not the framework authors. Not the marketplace runners. Not the developers shipping agents.
That's the gap.
We're Filling It
The Agent Smiths is a personality forge. We build characters for AI agents — complete kits (SOUL.md, VOICE.md, IDENTITY.md) that work with any framework that reads markdown.
Every personality we ship passes a battery of tests. The most important one: the Strip-the-Name Test. We take 10 responses. Remove the character name. Hand them to someone who's never seen the character. If they can't tell who's talking — we don't ship it.
We're giving away the catalog. Free. Because the craft speaks for itself. Download a personality. Drop it into your agent. See the difference between a character and a costume.
If you need something custom — a personality forged for your brand, your product, your specific agent — that's The Forge. Full workshop. Unlimited revisions. Every test passed before delivery.
Anthropic gave Claude a soul baked into the weights. Steinberger made it an editable file. 319,000 developers adopted it.
Now somebody needs to teach people what to put in it.