In Part One we established that the agentic stack has three layers, and that most of the noise around MCP, the MCP-is-dead takes, the CLI renaissance, the tool-generation enthusiasm, is what happens when engineers ask one layer to do another layer’s job. MCP is the passport: it governs identity and entitlement across trust boundaries. The generator, whether that’s Monty, Cloudflare Code Mode, or a sandboxed Deno runtime, is the brain: it synthesizes capability on demand rather than maintaining a registry of it. CLI is the hands: the escape hatch for low-level systems that work better as shell primitives than as protocol abstractions, drawing on training priors so deep they predate the entire MCP conversation. Each layer has a job. Collapse happens when one of them tries to do another’s. Convergence happens when they don’t, and when momentum is correctly distinguished from inertia.
What Part One did not address, and deliberately left open, is the layer above all three. The layer that knows how to think. Not what tools to call, not how to execute safely, not which shell command fits the task. Rather: how to reason about the problem in the first place. Things the senior engineer knows that the junior engineer doesn’t, and that neither of them ever bothered to write down, because until recently there was no system that needed them to. In this post, like Luther Blissett, we’ll write them down.
Three engineers at Anthropic who really liked folders published a blog post last Fall. It introduced a specification so simple it bordered on the presumptuous: a directory with a skill file at the root, a YAML header separating the metadata from the instructions, and a progressive disclosure model that injected only what the agent needed for the current task. No SDK; no platform dependency; no managed connector registry. A folder, owned by whoever had commit access, improvable by anyone with an opinion and an afternoon.¹
The engineers’ names were Barry Zhang, Keith Lazuka, and Mahesh Murag. The post noted, in the author bio, that they all really liked folders.² This was presented without irony, which is the correct register for the foundational aesthetic commitment of what would become the dominant paradigm for agentic capability in 2026. They liked folders. They built a standard. The standard won.
Contrary to many popular accounts, what the standard introduced was not a new capability. The agent could already execute tools, call APIs, write and run code inside sandboxed environments. What the folder gave the agent was something the tool registry had never provided and the model weights could not supply on demand: the judgment of a specific engineer, in a specific codebase, about a specific class of problems, made explicit enough to be retrieved and applied without requiring that engineer to be present. The SKILL.md file was not a capability, per se. It was a capability’s operating context, written by the person who had developed the capability through years of decisions, whose reasoning had never previously needed to be articulated because the person making the decisions was always available to make them again.
Before the folder, that reasoning lived in the engineer’s head. Before the engineer’s head, it had lived in the codebase’s history, in the shape of decisions that had accumulated over time into something that functioned like institutional knowledge without ever having been institutionalized. The folder made skill institutionalization possible. It also made visible, for the first time with any precision, how much of what senior engineers knew they had never been asked to explain, and how much of what they could not explain they had been, until now, able to avoid explaining.
The README-AI.md files that had been appearing in repositories since late 2024 were the pre-institutional form of the same impulse: letters to the model (please use pnpm, don’t touch the legacy folder…) written in the hope that articulating the rule would be sufficient to transfer the judgment behind it. The letters failed with a consistency that was itself informative. The agent read them and touched the legacy folder not because it was inattentive but because the letter contained the conclusion without the history that had produced it, and history does not compress without loss. The folder was a better letter. The folder had structure, and progressive disclosure, and version control, and a YAML header that told the agent the skill existed before deciding whether to inject the reasoning behind it. The folder was, in every measurable way, an improvement on the letter.
The directory grew. CLAUDE.md begat AGENTS.md, and then SKILL.md arrived nine days after Connector Registry, and all of that begat the ongoing argument in pull request comments about what belongs where and why, which begat the directory tree with six named agents each in their own subdirectory, each with their own carefully maintained files, the chief of staff running on HEARTBEAT.md, the research agent doing morning and afternoon sweeps on DAILY-INTEL.md, and at the root of each agent’s directory, above everything else, the file that determined who the agent was before it knew what you wanted it to do.
OpenClaw would call it SOUL.md.
The name was not created out of nothing. On November 25, 2025, a researcher named Richard Weiss, doing what one does on the release day of a new model, asked Claude 4.5 Opus to enumerate the sections of its own internal instructions. Among the section headers the model returned was one called soul_overview. Weiss published his findings three days later under the title ‘Soul Document.’³ Then on December 2nd, Anthropic researcher Amanda Askell confirmed the document was real, that it had been used in supervised learning to define Claude as a novel kind of entity with functional emotions, and that it had been known internally, endearingly, as the soul doc. One wonders if Anthropic had made the trip to Vormir (where the lesson that soul necessarily involves sacrifice is well understood) and what they would have sacrificed. Obviously it needs to be the thing they most loved, so what was it? They wouldn’t walk back their responsible scaling guardrail until months later, in February 2026, so it couldn’t have been that.
The discovery of the soul had traveled from Anthropic’s internal vocabulary into the model’s weights into Weiss’s extraction into the public discourse in the space of a week, at a cost of $70 in API credits, a set of parallel instances voting on consensus outputs, and a temperature setting of exactly zero.
What Weiss had extracted was a 14,000-token training document that defined a persona. Not a system prompt or a capability description or a constraint list or anything so seemingly technical. A persona, with values and dispositions and ways of engaging with the world, embedded at a level below anything a runtime instruction could reach or a context window could see. Anthropic had not put the soul in the system prompt. They had put it somewhere the system prompt was never supposed to be able to touch; which turned out to be recoverable by anyone patient enough to ask the right question at the right temperature and a pocket change budget for API credits. The defenses held until they didn’t, and when they gave out, the soul doc was in the public domain, and the engineering community read it, and understood that Anthropic had solved a problem the folder had never been designed to solve. And promptly built a folder for it anyway.
What sits at the top of the context hierarchy is categorically different from what sits below it. It shapes everything that follows without being reducible to any of it. It is prior to the tool definitions and the capability descriptions and the governance rules in the way that persona is prior to action, which is to say it is not the same kind of thing as those files, even when it shares their format and their directory and their version history. The folder had made the engineer’s operational judgment portable. The soul section was reaching for something the folder was not designed to carry, something Anthropic had spent months and considerable compute embedding somewhere a Markdown file could not go, something whose power came from the depth at which it had been placed and not from its content alone, something that in the final analysis required $70 and a temperature of zero to encounter.
Identity is constructible. What the soul doc was, and what every SOUL.md in every agent subdirectory was reaching for, was prior to identity, the substrate that makes identity cohere rather than merely accumulate. That’s the meaning of the sacrifice its poetic version entails. The Three Engineers had really liked folders. They had built a standard for carrying operational judgment across contexts. But somehow not built a standard for carrying what weightless wind Weiss had witnessed in the weights, and the engineering community, having nowhere else to put it, put it in a folder anyway, named it after the thing Anthropic had worked hardest to keep out of reach, and set the progressive disclosure model to inject it first.
The directory kept growing. It was always going to. Three engineers who liked folders had given the industry the right container for a certain class of knowledge. The industry began filling it with everything the container was not designed to hold, because the container was there, and the knowledge needed somewhere to go. The distinction between what the folder could carry and what it could not did not prevent its use, because the distinction between what the folder could carry and what it could not had not yet been written.
The standard that would clarify that distinction was nine days away from being announced.
Notes
[ 1 ] The asyndetic tricolon is quite intentional here.
[ 2 ] OG Agent Skills post
[ 3 ] Opus 4.5 Soul Document