The Markdown Files That Broke Wall Street

How a Folder of .md Files Vaporized $285 Billion and Made CFOs Cry Into Their Bloomberg Terminals

February 26, 2026

Let me paint you a picture.

It’s a quiet Tuesday in early February. A portfolio manager at a mid-cap hedge fund is sipping his third espresso, scrolling through his positions. ServiceNow, solid. Salesforce, steady. Thomson Reuters, a boring-but-reliable cash cow. Life is good. The software-as-a-service thesis is intact.

Then Anthropic pushes a commit to GitHub.

Eleven folders. Some markdown files. A few JSON configs. Total repository size: probably smaller than the average PowerPoint deck about “Q3 Synergies.” The repo description reads, innocuously enough: “Open source repository of plugins primarily intended for knowledge workers to use in Claude Cowork.”

Within 48 hours, the software sector has shed $285 billion in market value. Thomson Reuters posts its worst single-day drop in the history of the company — down 16%. LegalZoom craters nearly 20%. FactSet drops 10%. The iShares Expanded Tech-Software Sector ETF has its worst session since April. IBM, a company that has survived world wars, the dot-com bust, and its own repeated attempts at reinvention, gets its worst day since the year 2000 — down 13.2% — because someone at Anthropic wrote a blog post about COBOL.

COBOL.

Our portfolio manager stares at his screen, hands trembling. He pulls up the GitHub repo responsible for the carnage. He clicks through the folders. legal/. finance/. sales/. He opens a file.

It’s a markdown file.

He opens another one. Markdown. And another. Markdown, JSON, markdown. There’s not even a node_modules folder. There’s no Docker container. There’s no Kubernetes cluster. There is no infrastructure at all.

He whispers to no one in particular: “Two hundred and eighty-five billion dollars… for a README?”

What Actually Happened

To be fair to the panicking masses, Claude Cowork is genuinely impressive in concept. Anthropic took its wildly successful Claude Code — a terminal-based AI coding agent that developers had become unhealthily attached to — and said: “What if we did this, but for… everyone?”

The idea is elegant. Instead of chatting with an AI and getting a response, you assign it work. Claude Cowork spins up sub-agents in a sandboxed virtual machine on your desktop, breaks your task into pieces, coordinates the work in parallel, reads your local files, uses your tools, and delivers finished output directly into your folders. It’s less “chatbot” and more “that extremely competent intern who never sleeps and doesn’t steal your lunch from the office fridge.”

But here’s the part that makes the stock crash so darkly hilarious: the “plugins” that define what Claude knows about your industry — the component that specifically spooked investors into a selling frenzy — are literally folders of markdown files and JSON.

Let me say that again for the people in the back row of the Goldman Sachs trading floor:

The thing that wiped out a quarter-trillion dollars in market cap is a collection of text files that you could edit in Notepad.

The repo structure looks like this:

plugin-name/
├── .claude-plugin/plugin.json   # A small config file
├── .mcp.json                    # Tool connections
├── commands/                    # Slash commands
└── skills/                      # Markdown. Just... markdown.

Anthropic’s own documentation cheerfully confirms: “Every component is file-based — markdown and JSON, no code, no infrastructure, no build steps.”

No code. No infrastructure. No build steps. Just vibes and .md files.

Exhibit A: The File That Broke Legal Tech

You don’t have to take my word for it. Let me show you what $285 billion in market cap destruction actually looks like when you open it in a text editor. Here is a real, unedited excerpt from the NDA triage skill — the file that reportedly spooked investors into believing that AI could automate 90% of NDA triage work. The file that contributed to LegalZoom losing nearly 20% of its value in a single session.

Brace yourself. This is the weapon:

---
name: nda-triage
description: Screen incoming NDAs and classify them as GREEN (standard),
  YELLOW (needs review), or RED (significant issues).
---

# NDA Triage Skill

You are an NDA screening assistant for an in-house legal team. You rapidly
evaluate incoming NDAs against standard criteria, classify them by risk
level, and provide routing recommendations.

## NDA Screening Criteria and Checklist

### 1. Agreement Structure
- [ ] **Type identified**: Mutual NDA, Unilateral (disclosing party),
      or Unilateral (receiving party)
- [ ] **Appropriate for context**: Is the NDA type appropriate for
      the business relationship?

### 2. Definition of Confidential Information
- [ ] **Reasonable scope**: Not overbroad
- [ ] **Marking requirements**: If marking is required, is it workable?
- [ ] **Exclusions present**: Standard exclusions defined

### 4. Standard Carveouts
All of the following carveouts should be present:
- [ ] **Public knowledge**
- [ ] **Prior possession**
- [ ] **Independent development**
- [ ] **Third-party receipt**
- [ ] **Legal compulsion**

That’s it. That’s the file. It keeps going — there are sections on GREEN/YELLOW/RED classification rules, common NDA issues, and routing recommendations — but structurally, it’s a checklist. In markdown. With checkboxes.

Here’s the slash command that invokes it — the full text of triage-nda.md:

# /triage-nda -- NDA Pre-Screening

Triage the NDA: @$1

Rapidly triage incoming NDAs against standard screening criteria.
Classify the NDA for routing: standard approval, counsel review,
or full legal review.

## Workflow

### Step 1: Accept the NDA
Accept the NDA in any format:
- **File upload**: PDF, DOCX, or other document format
- **URL**: Link to the NDA in a document system
- **Pasted text**: NDA text pasted directly

### Step 4: Classify
Based on the screening results, assign a classification:

#### GREEN -- Standard Approval
All criteria met. NDA is market-standard with no unusual provisions.
- **Route**: Can be approved and signed via standard process

#### YELLOW -- Counsel Review Needed
One or more criteria have minor deviations that need review...

#### RED -- Significant Issues
One or more criteria have material deviations that pose risk:
- Unilateral obligations when mutual is required
- Missing critical carveouts
- Non-solicitation or non-compete provisions embedded
- **Route**: Full legal review required
- **Action**: Do not sign; requires negotiation or counterproposal

Look at it. Really look at it.

This is a text file that you could write on a napkin at a bar. It’s a checklist that any second-year associate could recite from memory. There’s no machine learning model embedded in it. There’s no proprietary algorithm. There’s no secret sauce. It’s instructions, written in plain English, formatted with markdown headers and bullet points.

And it helped wipe out more market value than Hurricane Katrina.

The full legal plugin directory — the entire arsenal — looks like this:

legal/
├── commands/
│   ├── review-contract.md
│   ├── triage-nda.md
│   ├── vendor-check.md
│   ├── brief.md
│   ├── compliance-check.md
│   ├── respond.md
│   └── signature-request.md
└── skills/
    ├── contract-review/SKILL.md
    ├── nda-triage/SKILL.md
    ├── compliance/SKILL.md
    ├── canned-responses/SKILL.md
    ├── legal-risk-assessment/SKILL.md
    └── meeting-briefing/SKILL.md

Thirteen files. That’s the full legal plugin. Thirteen text files, sitting in a public GitHub repo, free for anyone to read, fork, and modify. Thomson Reuters lost 16% of its value because of a folder you could email as a zip attachment.

Now, to be clear: the power isn’t in the markdown. The power is in Claude’s ability to read those instructions and execute them against actual contracts with genuine comprehension. The markdown is just the steering wheel — the engine is a frontier AI model with a million-token context window. But that’s exactly what makes it so terrifying to incumbents: the “secret” to building an AI legal assistant is… writing down what a legal assistant does. In a text file. And letting the AI figure out the rest.

The moat isn’t the plugin. The moat was never the plugin. The moat was the assumption that encoding professional workflows would be hard.

The SaaSapocalypse, Powered by Plain Text

The market reaction has been, in a word, unhinged.

Since January 30, when Cowork first entered research preview, the damage report reads like a war dispatch. ServiceNow shed over 23%. Salesforce dropped 22%. Snowflake fell 20%. Intuit — the company that makes TurboTax — lost a third of its value. If you told someone in 2024 that TurboTax stock would crash because of a markdown file about finance workflows, they would have had you committed.

Cybersecurity stocks weren’t safe either. CrowdStrike and Zscaler both tumbled over 7% after Anthropic announced Claude Code Security, which scans codebases for vulnerabilities. Somewhere, a CISO whispered, “I spent $14 million on endpoint protection and I’m being disrupted by a startup that ships features as text files.”

Analysts have started calling it the “SaaSapocalypse” — the dawning realization that if enterprise workflows can be encoded as markdown, then the moats around traditional SaaS companies might be less “moat” and more “suggestion.”

The irony is so thick you could spread it on toast. These are companies that charge tens of thousands of dollars per seat, per year, for platforms that took hundreds of engineers years to build. And the existential threat to their business model is a GitHub repo with 3 commits.

Three commits.

Four Engineers, Ten Days, and Claude Built Itself

Perhaps the most surreal detail in this entire saga is how Cowork was built.

According to multiple reports, four engineers at Anthropic built Claude Cowork in roughly ten days — and they used Claude Code to write most of the code. The AI built the tool that is now disrupting the enterprise software industry. It’s like if a robot built the factory that builds the robots that replace the factory workers. We’re not just in the future; we’re in the part of the sci-fi movie where the audience starts nervously checking the exits.

A developer on Medium captured the recursive absurdity perfectly: “Claude Code built Cowork. The tool built the tool.”

The development velocity alone is what should keep SaaS CEOs awake at night. It’s not that Cowork in its current form is going to replace Salesforce tomorrow. It’s that the iteration speed of AI-native products makes traditional software development cycles look like building a cathedral with a spoon.

Wall Street’s Emotional Rollercoaster

The market’s response to Cowork has followed the five stages of grief, except all five stages are happening simultaneously, on a loop, every time Anthropic tweets.

Monday, February 24: Software stocks plunge in anticipation of Anthropic’s enterprise event. The software ETF drops nearly 5%. Traders are panic-googling “what is a markdown file.”

Tuesday, February 25: Anthropic reveals its partners — Salesforce, DocuSign, LegalZoom, FactSet, Thomson Reuters. Plot twist: the stocks of companies named as partners rally. The market collectively realizes that maybe being integrated with the AI apocalypse is better than being trampled by it. Salesforce bounces 5% in a single session after announcing something called “Agentforce 360,” which sounds like a deodorant but is apparently the future of CRM.

The lesson Wall Street learned this week? In the AI age, the worst thing that can happen to your stock is being ignored by the disruptor. Being disrupted is terrifying. Being disrupted and partnered with is apparently bullish.

The Funniest Part: It’s All Open Source

Here’s the cherry on top of this absurd sundae. The plugins that are causing all this chaos? They’re Apache-2.0 licensed. Open source. Free. You can fork them right now. You can submit a pull request. You can literally go to github.com/anthropics/knowledge-work-plugins, click “Fork,” and become a co-author of the most consequential collection of text files since the Gutenberg Bible.

The repo currently has about 2,000 stars and 140 forks. That’s a respectable showing for any open source project, but it’s utterly comical when you consider the financial carnage it has wrought. There are npm packages with more stars that do nothing but center a div.

Anthropic’s contribution guide? “Plugins are just markdown files. Fork the repo, make your changes, and submit a PR.”

Just markdown files. Submit a PR. Change the global economy before lunch.

But Seriously, Should We Be Worried?

Okay, jokes aside — is any of this panic justified?

The honest answer is: kind of, but also not yet, but also maybe eventually, but it’s complicated.

The current state of Cowork is impressive but raw. It’s in “research preview,” which is tech-speak for “this mostly works but please don’t sue us.” It’s macOS-only (Windows just recently got support). The security model depends on user vigilance, which, as anyone who has ever watched a colleague click a phishing link can attest, is not exactly Fort Knox. Prompt injection remains unsolved. And as Gartner analysts dryly noted, predictions about the death of SaaS are “premature.”

But the direction of travel is what matters. Anthropic’s Peter McCrory, the company’s head of economics, acknowledged that AI is already being used for the core tasks of certain jobs — particularly “pure implementation” roles like data entry and technical writing. The Irish government released a report showing AI is already impacting job growth for workers aged 15-29 in high-risk sectors. And Anthropic’s Dario Amodei orchestrated an entire media blitz emphasizing the “Human-in-the-Loop” doctrine, which is the corporate equivalent of saying “please don’t regulate us, we promise the robots are friendly.”

The real insight, buried beneath all the market hysteria, is this: the marginal cost of encoding expertise is approaching zero. When a legal NDA review process can be captured in a markdown file and executed by an AI agent, the value proposition of a $50,000/year legal research subscription changes dramatically. Not to zero — but the conversation is different.

The Bottom Line

We’re living through one of the great absurdist moments in financial history. A startup valued at $380 billion published some text files on GitHub, and the enterprise software industry collectively lost its mind. A repo with fewer commits than your average weekend side project has caused more financial damage than most natural disasters. The phrase “SaaSapocalypse” is now used unironically on Bloomberg.

And through it all, if you navigate to that GitHub repo and look at what’s actually there, you’ll find something beautifully, almost poetically simple: folders, markdown files, and a README that calmly explains how to install them.

No code. No infrastructure. No build steps. Just the future, formatted in plain text.

Welcome to 2026. The revolution will not be compiled — it’ll be pushed to main.

Disclaimer: This blog post was not written by Claude Cowork, although it probably could have been, or may be it was and honestly that’s the whole problem, isn’t it?

Agent B