Skills, MCP and Plugins: Making a Method Reusable

Ask a general assistant to review an NDA against your firm's position and it hits two walls. It does not know your position, so it falls back on a generic checklist. And it cannot open the NDA, because the file is in your document system and it has no way in.

So you download the file, paste it in, and type out your position from memory. Then you do it again for the next NDA. Chapters 15 and 16 taught you to do that well. This chapter is about the part nobody mentions: doing it well by hand does not survive the tab you did it in. The bottleneck in legal AI is almost never the model. It is that the method is not written down anywhere.

Three add-ons remove those walls, and you do not need to build any of them to benefit. They show up in product settings and vendor pitches, and between them they explain most of what a legal AI tool can actually do with your work.

A skill is not a prompt

This distinction is worth being precise about, because the three things look identical in a text box and behave nothing alike.

• A prompt is what you type in one conversation. It works once. When the conversation ends it is gone, and next time you type it again, slightly differently.
• Standing instructions (a project's custom instructions, a system prompt) are saved, but they apply to every message whether they are relevant or not, so they compete for the same finite context windowThe total amount of text a model can have in front of it at once: the standing instructions, your attached documents, the conversation so far and your new message. Measured in tokens. as your actual documents.
• A skill is saved and is loaded only when the task matches it. You write the guidance once and never repeat it.

That last clause is the whole mechanism, and it is what makes installing dozens of skills possible rather than absurd. At the start of a conversation the assistant sees only each skill's name and one-line description, roughly a hundred tokens apiece. When your request matches a description, it opens that one skill and reads the instructions in full Anthropic · Agent Skills (opens in a new tab). Try it with four installed and one request.

The same rule applies inside the folder, which is why a skill has no practical size limit. A skill can be one sentence ("when you write to a client, sign off with the matter number") or a thousand lines carrying every clause of your NDA position and the reasoning behind each one. The published guidance is to keep the entry file short and move the detail into separate files it points at Anthropic · Skill authoring best practices (opens in a new tab). Those open only when a step needs them, and a program is run rather than read, so only its result enters the conversation.

Why MCP matters to a buyer, not just an engineer

MCP is an open protocol for connecting assistants to the systems that hold your data, published in November 2024 Anthropic · Introducing the Model Context Protocol, Nov 2024 (opens in a new tab) with an open specification MCP · Specification (opens in a new tab). That sounds like an engineering detail. It is a procurement question, and the arithmetic shows why.

Before a standard, every integration was bespoke: your document system to vendor A, then the same work again for vendor B. Most pairings were therefore never built at all. And the cost of the ones that were is not the engineering. It is that you cannot leave. Switching assistant means rebuilding every connection, so the first vendor through the door keeps you by inertia rather than by merit.

A connector written to an open protocol is one you can point at a different model next year. That is the whole of the argument, and it is the question to put to a vendor: is your integration layer a standard, or is it yours? The red flags in How to Evaluate a Legal AI Vendor all point at the same thing.

Two cautions on connectors specifically. Whatever the assistant reads through one goes to the model's provider exactly as if you had typed it, so Privilege & Confidentiality with AI applies in full. And text arriving through one, an email or a document from the other side, can carry hidden instructions aimed at the assistant rather than at you. A well-built connector uses your own access rights: if you cannot open a matter in the document system, neither can the assistant. Ask whether that is true of the one you are being sold.

How the pieces fit

Back to the Acme NDA, with a contracts plugin installed. You type "review the Acme NDA against our position." The assistant matches the request to the NDA skill and reads your checklist. It uses the document system connector to find and open the file. The model then does what chapter 13 described: it reads, compares and writes, in the format your skill asked for. You review the result.

The skill supplied the method. The connector supplied the file. The plugin delivered both. The model did the reading and the writing, and it was the same model all along.

That last sentence is the buyer's lesson, and it is worth sitting with: much of what makes an AI tool "legal" is skills and connectors arranged around a general model. Which is good news if you are choosing, because it means the differentiator is inspectable. Ask to see the skills. They are plain text.

Written once, then reused

Writing a method down costs more than doing the task once. That is the honest objection, and it decides which tasks are worth the trouble. For a task you do once a year the procedure takes longer than the work, so do the work. For a task you do every week the cost is paid once and then stops.

Shared or private, and why that is a strategy question

Some methods have no competitive content. A procedure for checking an NDA against standard positions, or a classification checklist for a regulatory notice, does not need rebuilding inside every legal team in the market; sharing it costs you nothing and raises the floor. Others are the practice itself: your fallback positions, your risk appetite, what your firm concedes and what it never does. Those stay yours, and The Data Myth explains why they, and not your document archive, are what makes a firm defensible.

Either way the method belongs in a structured, reusable form rather than in somebody's head, which is also the only version of it you can measure, improve and prove. That last point is the next chapter.