The Trillion-Dollar Markdown File: Unpacking the 'SaaSpocalypse'
Published on
The Secret I Thought Everyone Knew
For the better part of a year now, I have been using a Large Language Model to do initial passes on contract reviews.
It wasn't a particularly sophisticated setup. With a nicely tailored system prompt, a few guardrails, and some initial legal insight from a few colleagues to refine the instructions, it worked brilliantly. It allowed my team to do rapid-fire review cycles both before and after the formal (and expensive) human legal review. It caught obvious red flags, standardized clauses, and saved everyone a massive amount of time.
So, on January 30th, when Anthropic announced they were releasing 11 open-source "Cowork plugins" for Claude - including a specific "Legal" plugin for contract review, compliance tracking, and NDA triage - my reaction was a mild nod of approval. I thought, “That’s nice. They’ve packaged it up so it's more accessible to people who don't want to write their own system prompts.”
I went back to my coffee. The global financial markets, however, absolutely lost their minds.
The Knee-Jerk Wipeout
What followed was dubbed by the financial press as the "SaaSpocalypse."
In the immediate aftermath of Anthropic's announcement, investors panicked. The assumption seemed to be that this new "Legal AI" was going to instantly render the entire incumbent legal-tech and SaaS industry obsolete.
The numbers are genuinely staggering. Thomson Reuters (operators of the legal database Westlaw) saw their stock plummet by nearly 18%. RELX, the parent company of LexisNexis, suffered its steepest single-day decline since 1988, dropping over 14%. Wolters Kluwer dropped 13%, and LegalZoom took a 20% hit.
In just five trading days, a combined $285 billion was wiped off the market capitalization of these software and data service providers. A week later, on February 6th, when Anthropic launched Claude Opus 4.6 with multi-agent coordination, the cumulative market wipeout across the broader software sector pushed toward the $1 trillion mark.
It's a catchy number for a headline, even if it took a few extra days to get there. But all of this carnage was triggered by the release of a plugin. Which made me wonder: what exactly did everyone think Anthropic had built?
Peeking Behind the Curtain
I decided to go digging. Anthropic made the plugins open-source, so I navigated to their GitHub repository to look at the underlying code for this market-destroying legal juggernaut.
Here is the exact file on GitHub: anthropics/knowledge-work-plugins/blob/main/legal/skills/review-contract/SKILL.md.
Given the sheer scale of the market panic, I can only assume that institutional investors and analysts expected to uncover a highly complex integration pipeline, a proprietary legal reasoning engine, or a custom RAG architecture fine-tuned on centuries of obscure case law.
I, on the other hand, found exactly what I expected to find: A Markdown file.
It is literally just a text document containing system instructions telling Claude how to approach reading a legal document. It dictates the tone, specifies the output format compliance teams like to see, and provides a structured workflow. That is it. There is no magic legal brain. It is just Claude being Claude, wrapped in a very well-written set of instructions.
Where the Moat Actually Lives
This staggering disconnect between the underlying technology and the market's reaction highlighted two glaring realities for me.
1. People really do not understand how these tools work. The broader public - and apparently, Wall Street analysts managing billions of dollars - see a "Legal AI Plugin" and assume it is a distinct, highly advanced piece of software fundamentally different from the chatbot they use to write emails. They don't realize that the intelligence isn't in the plugin; the intelligence is the base model, and the plugin is just a set of polite directions.
2. The "Proprietary" SaaS grift is being exposed. How many organizations over the last two years have built so-called "proprietary" legal or financial AI tools that are functionally identical to this Markdown file? There is an entire cottage industry of startups commanding massive valuations for software that simply takes a user's document, slaps a hidden text prompt on it, feeds it to an API, and returns the result in a slick UI.
The reality of the AI industry is that the true value - and the astronomical expense - sits entirely at the foundation layer. The real moat is the billion-dollar clusters of GPUs running 24/7, the exabytes of curated training data, and the armies of researchers tweaking neural network architectures. The heavy lifting is done by the models themselves. The wrapper companies simply built a nice steering wheel for someone else's spaceship. Anthropic publishing those prompts for free wasn't a sudden, terrifying leap in AI capability; it was just the destruction of an artificial moat.
The Four-Minute Prototype
There is an extra layer of irony to this whole saga.
Given how AI development works today, it is highly probable that the Anthropic team didn't even write that Markdown file from scratch. They likely just asked Claude to draft the optimal system prompt for a legal review tool.
Think about that for a second. A trillion dollars of market capitalization was effectively destabilized by a text file that likely took an AI about four seconds to generate. Has so much imaginary wealth ever been wiped out by something so incredibly simple? (This isn't to mock the staggering underlying complexity of Claude or the ecosystem Anthropic has built, but the sheer leverage of it is undeniably amusing).
To prove just how accessible this leverage is, I ran a quick experiment at dinner the other night. After a couple of glasses of wine, using only my phone and my admittedly limited formal legal expertise, I showed a friend how to create a similar tool. Using an LLM to help me write the prompt, I spun up a custom, shareable "proprietary legal assistant" that did the exact same thing as the market-crashing plugin.
From idea to a fully functional, ready-to-use tool took roughly four minutes.
The Bipolar Market
There is a deep, mildly comedic irony to all of this. If you follow the financial news, this pendulum swing is a fairly regular occurrence.
For months, you will find extensive op-eds declaring that Generative AI is a massive, overvalued bubble. Pundits will boldly claim that the technology is plateauing, that it's just a glorified autocomplete, and that the tech giants are burning capital on a hype train that will never yield real enterprise value.
Then, suddenly, a tech company uploads a 3-kilobyte text file to GitHub, and those exact same markets immediately enter into a cold sweat, sell off a trillion dollars worth of established software stocks, and run screaming for the hills because the AI is going to take over everything and wipe out all human enterprise.
It seems we are currently trapped between treating AI as a complete disappointment and a literal god. As for me? I'm going to let Claude read my NDAs so I have more free time to use a sprawling cluster of cloud GPUs to flawlessly erase a discarded kebab wrapper from a photo of a Golden Retriever.