Insights

Claude for Legal (May 2026 release)

On May 12, 2026, Anthropic shipped the "Claude for Legal" marketplace plugins covering a wide array of legal practice disciplines and workflows — plus 20+ MCP connectors linking Claude to Ironclad, iManage, NetDocuments, CoCounsel, Westlaw, Datasite, and more. These aren't products to license. They're open-source, free via GitHub, and closer to a package manager than a SaaS vendor. Each plugin contains a setup interview that captures your org's playbook, escalation chain, risk calibration, and house style. This summary covers plugin architecture, four selected plugins (commercial, IP, litigation, employment), and near/long-term strategy considerations for legal teams evaluating adoption.

Business Considerations Following the Stainless Acquisition

Anthropic acquired Stainless this week, and most of the (limited) takes I'm seeing are missing what actually changed. Also, with Claude for Legal dropping last week (more on that later), Karpathy, all the IPO talk, this story was easy to miss. Most headlines focused on the customer list: OpenAI, Google, and Meta reportedly all used Stainless to generate their SDKs. Interesting, but not the real story. The real story is the generator itself, and the clear signal that models are no longer the moat. Anthropic just bought the factory that converts API definitions into agent-ready interfaces at scale. That has real implications for anyone building, buying, or betting on MCP servers right now — and a few of them are not obvious. I put together a short deck walking through what I think shifts from here: where the moat moves, what becomes a strategic asset, where the lock-in risk actually lives, and the KPI I'd be tracking in 12 months.

Extensbile Cowork Plugins Using plugin2mcp

Re: Anthropic Claude Cowork plugins - it is looking like the masses are starting to realize what these plugins are (commands+instructions) and what they are not (enterprise SaaS replacements), which is good. Even Wall Street has moderately corrected its wild over-correction that caused $trillions of paper losses last week. Regardless, I expect Cowork to gain huge momentum this year, and with that I also expect plugins to become a popular solution framework for repeatable AI 'skills' that are easy to use. Moreover, I expect the uptick in plugins will only increase the demand for MCP tools that run complex workloads outside of Cowork. For anyone interested in that last bit - specifically the development of said MCP tools - this post is for you...

Claude Cowork Plugin Extensibility

There's been a ton of hype-chatter since I posted about the new Anthropic Claude Cowork plugins on Sunday, and rightfully so. Not because it turns law or technology on its head, but because it demonstrates a substantial acceleration towards putting powerful SME-specific tooling in the hands of working professionals for less than the cost of Netflix (and a lot more useful). There's also, unsurprisingly, been a fair amount of gatekeeping coming from hashtag#legaltech vendors, and that's also rightfully so. Anyone with experience building or procuring enterprise software should be well aware of the massive differences between single-user productivity enablers and centralized enterprise solutions. (think: Notepad++ vs Office365). But not all legal tech is truly enterprise, even if it's pitched that way. And some of the rebukes I'm seeing are indirectly exposing that truth, because they are grounded in value propositions that are easily replicated. Here is a real but [REDACTED] example: Contract AI employee says Cowork's Legal plugin isn't that great, because it just dumps an analysis in chat rather than integrating feedback into Word where the user really works. While this is a technically true statement, it's a weak argument and it's one that caught me in an opportunistic mood...

Claude Cowork Plugin Launch

I had some time this weekend to experiment with the freshly released Cowork plugins published by Anthropic that cover ~10 industry/job functions like marketing, finance, and more notably to me, legal and product management. Where Cowork brings a pre-loaded UI-driven agentic experience for non-technical users (e.g. Claude Code with UI), plugins bring pre-configured skill specialization to those same users. My initial takeaway is that this is the next inevitable step towards making LLMs usable, customizable, and accessible to millions of users. To be clear, these plugins are markdown instruction files that anyone could create and use like traditional model instructions. For each plugin, there's one skill-md and one commands-md. Users can open them, edit them, copy them... and I'm sure that plenty will, although that's not the intent. For example, the Legal plugin prefers to be left alone and instead asks the user to customize a separate playbook-md that gets referenced by the skill. Regardless of how users (or entire departments/organizations) choose to customize these or not, the simple act of publishing them as kickstarters for specialized use cases is powerful...

CPC Classifcation Agent

As part of a larger solution, I built an agentic workflow to perform CPC classifications for new invention disclosures (or reclassification of existing patents). There are probably several ways to architect such a solution, but this one is a good mix of secure, fast, comprehensive, and scalable. It uses a step-through process with a series of handoffs and inline validations at each step to prevent hallucinations and improve accuracy. It also provides detailed reasoning to provide transparency, which also allows the user to nudge it in a slightly different direction if need be. For those unfamiliar with the CPC, it's a patent taxonomy that is used to group patents together into similar technology segments. In an effort to globally standardize patent classifications, the USPTO moved to this system ~10 yrs ago in collaboration with the European Patent Office. The ability to rapidly classify patents/disclosures is of course useful for examiners, but it's also an invaluable tool for attorneys and inventors who often need to look through prior art (past publications of similar technology)...

AI Is Not Upending The Law Firm Landscape

My LinkedIn feed is suddenly being dominated by folks proclaiming massive, macro-level changes in law firm viability at various levels due to AI. Recently I've seen a multitude of (contradictory) predictions that AI will: - Siphon substantial amounts of traditionally complex work away from large law firms - Render boutique & solo practices redundant due to increased DIY - Cause mid-size firms to collapse due to being stuck between agility and purchasing power All of these statements are wildly unfounded. AI is a great tool. It is simultaneously a service delivery enhancer, an internal optimizer, and an innovation incubator. No doubt we are at a pivotal inflection point that hinges on technology adoption, but we've been here before. We aren't likely to see an earth shattering paradigm shift, and here's why:...

Lost in the Middle: A Lesson in Failing AI Agents and Backwards Architecture

After a couple weeks of working on the next (yet-unreleased) vibe coding side project, I finally hit a wall. One of my agents - designed to handle complex, multi-step analytical tasks related to patent analysis - seemingly decided it was just going to lie its way into production and hope nobody noticed. It would perform admirably for a while on fresh tasks. But as the context grew, it would inevitably begin to hallucinate, or just fail outright (which is way better but still bad). Not only would it hallucinate content, it would actively find ways around validation checks and guardrails in order to provide that content. The fabrications would at times be so subtle, only by looking at logs could one tell it was faked. But hallucinations are not what this article is about. To the contrary, one thing I've learned is that hallucinations are usually just a symptom - byproducts of broken information pathways or otherwise bad system design. It was that realization, or hunch, or maybe just unaware osmosis finally sinking in, that finally got things working reliably...

Hello World

index.query({ filter: { jurisdiction: 'hello' } })