AI Skills for Legal Agents: Vendor Library vs Your Own

By Jude Lee · October 9, 2026 · Custom

Lawyers and a paralegal reviewing an AI-drafted document checklist on a laptop in a firm conference room

What a skill actually is, without the marketing

Strip the branding and a skill is a document. It contains instructions (“when summarizing a deposition, produce a chronology table with page:line cites, then a credibility-issues section”), usually an example of good output, sometimes a checklist the model must complete, and sometimes a pointer to a tool it may call. The AI assistant loads that document when the task matches, and follows it. Anthropic publishes documentation for Agent Skills in Claude that works this way — a folder with a markdown instruction file plus optional supporting files. Other platforms wrap the same idea in a UI and call it a workflow, a playbook, or a template.

Why this matters more than it sounds: the difference between an AI that’s a toy and an AI that’s infrastructure is repeatability. A prompt typed fresh each time produces a different answer each time. A skill produces the same shape of answer from every associate, every matter, every Tuesday. That’s what makes review possible — and review is the only thing that makes AI output safe to put your name on.

A prompt is a conversation. A skill is a policy. Firms get burned when they treat the first as if it were the second.

Rent the skills or write them

Vendor skill library

Pre-built tasks shipped inside a legal AI platform — NDA review, chronology, clause comparison, research memo. Fast to start, maintained by the vendor, tuned to general market practice. Two questions to settle before you commit, because they decide the trade-off: pricing posture — is the library included in the seat price, or is it an add-on tier or per-task charge on top of seats? And instruction transparency — can you read, export, or edit the underlying prompt that produced the output, or is it a black box you must trust? Ask both in writing during the trial. Also assume the skills stay behind if you switch vendors.

Skills you write yourself

Plain instruction files you author and version — ideally in a repo or your document management system. Slower to start (you have to write down what “good” looks like, which most firms have never done). But they encode your templates, your jurisdiction’s quirks, your partner’s pet peeves. You can read every line, so review and supervision are straightforward. They’re portable across assistants, and they become a genuine firm asset that survives the next vendor cycle.

The third path: skills plus MCP

The Model Context Protocol is an open standard for giving an AI assistant governed access to your actual systems. A skill tells the agent how to do the job; an MCP connection gives it the matter file, the Clio contacts, the NetDocuments folder it needs to do the job on real data instead of pasted text. A skill without data access drafts in a vacuum. If you want the agent to read and write where your work already lives, see our walkthrough on connecting Claude to Clio with MCP and the deeper piece on building a custom MCP server over your matter data.

Which slice of your work deserves a written skill

The old 80/20 heuristic for lawyers — that a minority of your matters, clients, or task types drives most of your revenue and most of your headaches — is a useful filter here. You do not need hundreds of skills. You need skills for the handful of tasks that are (a) high-volume in your practice, (b) format-driven, and (c) currently done slightly differently by each person who touches them.

For a small plaintiff’s firm that might be medical chronologies, demand letters, and records requests. For a transactional shop: lease abstraction, closing checklists, and diligence summaries. For a family law practice: financial disclosure assembly and asset schedules. Everything else can stay as ad-hoc prompting or stay human.

5–12

Planning assumption: roughly how many skills cover a small firm's repeat work — count your own recurring task types, not ours

Illustrative planning range, not a measured benchmark

1–3 hrs

Planning assumption: time to draft a first skill if you already have an output you consider 'good'

Planning estimate — adapt to your own workflow

What firms are actually running today

Practitioner forums split into two camps: lawyers getting real leverage on drafting and review, and lawyers who tried a demo, got a confident wrong answer, and wrote the whole category off. Both are reporting accurately on different things.

In practice, the AI systems small and mid-size firms run cluster into four groups as of 2026. First, assistants — Claude, ChatGPT, Copilot — used for drafting, summarizing, and thinking out loud. Second, features bolted into practice management: Clio Duo, Smokeball Archie, MyCase’s AI tooling, mostly doing summaries, time-entry suggestions, and document search. Third, specialist legal platforms — CoCounsel, Harvey, Spellbook, Everlaw — where the skill library concept is most developed. Fourth, custom agents connected to firm systems via MCP or API, which is where firms go when the first three can’t touch their actual bottleneck. We compared that last trade-off in detail in Harvey vs CoCounsel vs a custom AI agent.

The most common real-world uses are unglamorous and that’s the point: first-draft correspondence, document review against a defined rubric, chronology building, discovery triage, intake qualification, and turning messy PDFs into structured data.

A sane build order

Pick one task you do weekly and hate

Not the most impressive one — the most repetitive one. Volume is what makes a skill pay for itself.
Write the gold standard first

Produce or find one output you’d be happy to send to a client. The skill’s job is to reproduce that, not to be creative.
Draft the skill as a plain file

Instructions, required output structure, a worked example, and an explicit “stop and flag” list of situations where it must defer to a human.
Run it on five closed matters

Closed files give you ground truth. Score the output against what the firm actually filed. Fix the skill, not the prompt.
Connect data only when the skill is stable

Wire it to your DMS or practice management via MCP once the reasoning is reliable — and start read-only. See our breakdown of read-only vs write access for firm AI agents.
Assign an owner and a review date

A skill with no owner rots. Someone re-reads it quarterly, or when a rule, form, or template changes.

Modeling what this is actually worth

Skip the vendor ROI headline and do your own arithmetic. For one skill: estimate the task’s current time, the realistic post-skill time including human review, multiply the difference by frequency, then by the rate of whoever currently does it.

(current hours − assisted hours) × tasks per month × billable or cost rate = monthly value

Plug in your own numbers. Three things change the answer more than the tool choice: whether recovered hours get reallocated to billable or business-development work (otherwise you’ve bought slack, not revenue), whether the skill captures time that was previously written off, and whether it prevents a specific error class you’ve actually paid for.

Lawyers chasing a larger personal book rarely get there by typing faster. They get there through rate, leverage, and matter mix — and skills are a leverage play: they let a smaller team handle more matters at a consistent quality bar. That only works if the recovered capacity is deliberately pointed at higher-value work. Our automation ROI calculator walkthrough steps through the full model including the reallocation assumption.

Where skills break

They break on judgment calls, on novel fact patterns, and on anything requiring knowledge the skill’s author didn’t anticipate. They break quietly when a form or local rule changes and nobody updates the file. And they produce confident, well-formatted, wrong output — which is more dangerous than obviously bad output because it passes the skim test.

The honest bottom line: start with a vendor library if one exists for your practice area and you want results this month. Write your own skills for the handful of tasks that define your firm’s work product. And be willing to conclude, for some workflows, that a deterministic template or a plain checklist beats an agent entirely. Skills earn their keep where the rubric is clear and the volume is real; they earn nothing where every matter is genuinely bespoke.

Where is your firm losing billable hours?

Get a free automation audit: we map your intake-to-invoice workflow and show you exactly what's worth automating — before you spend a dollar.

Get a free automation audit