Embed the Librarian Intelligence

Reading Time: 7 minutes

I am a little bit fascinated by the Anthropic-initiated Skills concept. In my mind’s eye, it fixes two of the significant problems with professional use of agentic AI (and generative AI): lack of expertise and lack of re-usability. The gap that I am unsure how to cross is moving from theory to application.

One of the challenges of teaching people how to gather information is dealing with the fact-specific nature of that gathering. We fall back on literacy as a baseline. The assumption being that we can teach someone how to fish, and they can adapt that knowledge to whatever body of water they happen to be working in.

That’s not entirely true. Law librarians most often will teach people how to use products from companies like Thomson Reuters or RELX or Alphabet. We stray from teaching universal approaches in order to move law students and researchers further along from literacy even though mastery is unlikely.

The goal up until now is for the learner to become the learned, for them to develop expertise through repetition and practice and increased understanding. This assumes the learner is doing the work. The change wrought by artificial intelligence is that that learning can be avoided. Organizations like law firms see artificial intelligence as having already achieved that learning. Meanwhile, law students, new graduates, and the public may be discouraged from engaging in the learning, with AI as a shortcut.

In some ways, it is the same issue that librarians have had ever since the emergence of internet web search. When people can ask an answer machine, they may disintermediate an information professional who had filled that role previously. It is not the librarian’s only role but web search narrowed the field of questions we answer and the types of reference materials we might collect.

The shortcoming with literacy is that, shorn of practice and repetition, it is unable to lead to expertise. When an AI is asked for information, the finding and retrieval is delegated. When an agentic AI is performing the task, that delegation occurs in an automated mode at an even greater distance from the potential learner.

Agentic Skills

One of the artificial intelligence companies created a “standard” called Skills. Anthropic designed it for people working with their product called Claude and released a specification that may provide a foundation for others to mimic. It’s a dumb name, almost as dumb as naming your platform X. I’m going to capitalize it for this post but hopefully it will spread and someone will call it something smarter.

The idea is that you can create templates that capture some of the know-how of more expert users. These Skills resources can then be exposed to AI agents who can determine whether to use them and, if they invoke them, will follow the deeper knowledge embedded in them. Anthropic has a GitHub site for Skills and defines them this way on the site:

This re-post over on O’Reilly’s Radar is a great explainer for what agent skills are from the perspective of someone using them. It notes that agentic AI is focused on completing tasks and not on the steps that a more experienced developer might take getting to the same result.

It is not a great leap to go from contemplating a “senior engineer” to “experienced librarian”. We know what we know. We cannot always impart everything we know to learners as we engage in developing information literacy. An expert researcher may have a suitcase full of tricks, any one of which will only be pulled out in a very specific context. Unless the limited options to teach and share skills touch on that context, they may never get an airing.

A Skill has a couple of components that will make sense to a librarian. The basic file—Skill.md (md for Markdown)—contains a metadata section and then the definition. It’s a bit like a catalog record or a web page. There is underlying metadata that describes the object that is ingested by the AI agent. If it decides the Skill is relevant based on the metadata, it then explores the rest of the object record (the Skill.md file).

If you’re curious about all of this, Anthropic has partnered with Skilljar on some very simple tutorials. A lot of it will be bafflegab to people who do not develop software with Claude but you can put the videos on 2x speed, read the text, and get a good general understanding. I took one session and that was plenty for me.

I found it more interesting to read through some of the samples. This really sparked my interest. For example, Anthropic has a sample Skill called PPTX for creating Microsoft PowerPoint slide decks. The creation involves design instructions, color branding specifications (in HEX), visual guidance, copy editing steps, and so on.

Note that the Anthropic training says to limit a Skill.md to 500 words and the PPTX Skill is 1200. It makes me think this is an idea that hasn’t fully been baked. Or the bean counters don’t want Skill.md files to become unwieldy and costly since, once they’re invoked, they use resources that have a cost. Also, the PPTX file has an alert that the agent should use “subagents — even for 2-3 slides. You’ve been staring at the code and will see what you expect”, as if the agent is a person. Nevermind that researchers think that being rude will actually get you better AI results.

A Basic Legal Research Skill

It made me wonder what a basic legal research Skill might look like. If I was going to boil down my process for searching for case law on a specific topic, what would go into that process? At the end of the day, Skill.md files look like process documents to me. You provide an agentic AI with the raw input (“what have Illinois courts held in disputes involving property owners where a tree extends across a property line and causes damage?”) and it runs off through the process.

So what might that look like in a Skill.md file?

• You might want to start with a section that ensures it is focusing on the correct jurisdiction. For example, explain which courts are binding on others or unusual names for courts or tribunals that might not be clear.
• A verification step that addresses hallucinations: test for citations, test for verbatim existence of text.
• A note up step that determines whether cases are good law.
• A step that puts the case citations into a Bluebook format or some other structure preparatory to being used in a document.
• A step that ensures that unpublished opinions are recognized as such and included or excluded based on local court rules.
• Guidance that prioritized types of resource (free v. commercial, in-plan v. out, a case law search-and-filter v. repeated case law searches when they are billed differently) to be used.

And so on. The Skill.md file could incorporate the expertise of a law librarian so that a lawyer who is doing a case law search would leverage that expertise even if a law librarian wasn’t involved. As the legal research tool set changes, the Skill resources would be updated (a new Python script to adapt to Bluebook changes when a new edition comes out, for example) to take that into account.

One concept I love is that this could be a template. As the Anthropic training showed, Skills can be shared across a team. Each lawyer would not need to create their own. A team of librarians could create and manage Skills to focus on specific types of research—cases, statutes, regulations, tribunals, boards, secondary sources, court rules—that an agentic AI could invoke depending on the search initiated by the lawyer.

Rubber, Meet Road

This is where I get stuck. I can see, from the Anthropic training, how this would work in a development environment. I do not see it from a legal professional environment and so there must be pieces missing for me. I think it’s the need to connect the agents with the legal research platforms. [Seriously, if I am getting any of this wrong, comments are open below]

Bob Ambrogi had a good piece about Anthropic and access to justice recently. It made me finally try to dig a bit further into how the pieces connect. It sounds like the model context protocol (MCP) is the missing ingredient. Thomson Reuters has recently announced an MCP connection with Anthropic, and Anthropic has posted about a bunch of other MCP “connectors” it has developed with law-related services. This post about building a simple agentic AI weather app using MCP has some clear graphics that explain how the process flows.

Frankly, it all starts to sound a lot like APIs all over again. Sure, sure, I get it, they’re not the same. But that’s where my brain went when I started reading about them. Also, I get how an agent would get from the left of that chart, above, using Claude Desktop, to the right side to an app like Thomson Reuters Westlaw, but I’m not sure why the connector has to be with Thomson Reuters CoCounsel, which would be on the … left side? The inclusion of CourtListener as an MCP connector for Anthropic is much clearer to me (and they explicitly say the “free new tool uses your API access on the CourtListener platform”).

I’d love to hear how a law firm deployed Skills. I also wish I had an environment where I could test them out more but the reality is the cost and complexity for experimentation remains outside of my reach. As Bob Ambrogi pointed out:

MCP is an open standard so it shouldn’t require a researcher to adopt a specific vendor. However, we may not see law-related connectors turned on for every AI agent platform. I’m assuming each one will require licensing or access management at the very least. On top of that, Skills would need to become just as common (and renamed).

Hopefully, over time, opportunities to try out Skills—if I have CoCounsel or Lexis Protege’, couldn’t they be exposed as part of my organization’s legal research interface; we can already share documents—will broaden. I think there is an interesting opportunity for experts like law librarians to pull up a chair to a research and retrieval table in a way that we may not have been able to for the last 20 years.