AI Skills for Law Firms: A Playbook That Runs the Same Way on Every Matter

An AI skill is a prompt written down so it runs the same way on every matter. It fixes the steps, the order, the guardrails and the output format in a folder, so the review a senior lawyer does in their head becomes something the practice can rely on, repeat and evidence. A demo convinces one partner. A skill is what convinces the committee.

Every legal team we work with has someone quietly brilliant with Claude. The work is often better than the process it replaces. Then it hits the same wall in every practice: which matter did the documents come from, would the review reach the same conclusions if it ran again next month, did anything privileged leave the circle, and who is supervising. A prompt living in one person's chat history cannot answer any of that. That gap is not a model problem. It is an engineering problem, and it is a solved one.

What separates a clever prompt from a legal AI skill

Three things move a prompt from personal cleverness to something a practice can run:

It is written down. The steps, the order and the guardrails sit in a file rather than in one associate's head.
It is fed from source. The document management system and the contract repository, not a folder somebody downloaded. That is the subject of our guide on connecting AI to your DMS and docket without losing privilege.
It leaves a trail. What was read, what was concluded and what a person approved, recorded as it happened rather than reconstructed later.

A long prompt pasted in fresh each time drifts. Slightly different steps, slightly different checks, until two reviews are no longer comparable. In a practice that reviews the same agreement type forty times a year, drift is what quietly destroys consistency.

The anatomy of a skill folder

A skill is a folder of markdown, with no build step. If your team can write a precedent note, your team can write a skill. The parts that matter:

SKILL.md carries what to do, in what order, what to refuse and what to escalate. This is the control. Version it.
CLAUDE.md is your playbook: positions, fallbacks, thresholds, house format, who approves what. Nobody else has yours. It is the differentiator.
references/ holds standing context: precedent language, definitions, jurisdiction notes, so the model does not guess at house convention.
CHECKS.md is our addition. It holds the assertions that must pass before output is shown: every position cited, every deviation flagged. It turns a draft into a reviewable one.

Skills are cheap to install because they load in stages. At session start only the name and description go into context. The full SKILL.md loads when a request matches its description, and the reference files load only when the body calls for them. You can install many skills without cost, but the description is the routing key: a skill that never fires looks identical to one you do not have, and nothing reports the miss. Name the phrases your team actually types.

Two guardrails that are specific to law

Both belong in the workflow rather than in training, because the Friday afternoon is exactly when a busy lawyer skips a check that a good one would run:

Matter isolation. Each matter gets its own workspace. The rule is stated plainly in the skill: never read another matter's files unless cross-matter context has been deliberately switched on. That is an information barrier expressed as a file path. It is not a substitute for your conflicts system, but it is the difference between an assistant that respects the barrier and one that has never heard of it.
The destination check. Before output is produced, the skill asks where it is going. A public channel, a company-wide list, a counterparty or a vendor can each waive protection. When the destination looks outside the circle, it offers a privileged version for legal, a sanitised version for the wider audience, or both. What it will not do is stamp a privileged header on something you are about to paste where the header will not protect it.

Six skills to start with

Adapted from the open legal library, each works on documents you upload and gets sharper reading from a connected system:

/tabular-review builds a diligence grid, one row per document, one column per data point, every cell carrying the verbatim quote and its location.
/dpa-review works out whether you are controller or processor before it starts, then applies the right half of your playbook and checks the result against your published privacy policy.
/privilege-log-review makes the obvious calls, flags the hard ones, and refuses to decide the close ones. That restraint is the design.
/is-this-a-problem answers the four-o'clock-on-a-Thursday triage question with one of three verdicts: fine, needs a look, or hold.
reg-feed-watcher checks the feeds you name on a schedule, filters by your materiality threshold, and drafts a digest.
/stakeholder-summary turns a completed review into the two-minute answer a business owner will actually read.

Every cell, every digest item and every verdict is a lead, not a finding. The grid tells you where to look. A lawyer still reads the clause. That distinction is the whole professional obligation, and the published skills say it repeatedly because they are right to.

The playbook is the product

The six skills are deliberately generic. What makes them yours is the file they read before they start: your positions, your fallbacks, your escalation thresholds, the two clauses you never accept and the one you always trade. In most practices that file is undocumented, held by three or four senior people, and re-explained to every new joiner. Writing it down is work you should be doing anyway.

Do not compose it from scratch. Take a real review you are doing this week, work through it with Claude until the output matches what you would have produced, then ask it to turn what you just did into a skill and read the folder it drafts. That reading is where you find out what you never actually specified.

Before it goes near a live matter, a named partner or GC should own it, it should live in version control so a change to a control is reviewable, it should have run against a closed matter with the output compared to what was actually done, and it should declare what it could not check rather than staying quiet. The next step, packaging that folder so a colleague can install it, is covered in our guide on plugins and agents for legal practice. For the wider argument, start with the AI guide for legal practice.

The useful test is your insurer's question: if they asked how a conclusion in last quarter's diligence report was reached, could you answer without asking the associate who ran it? If the answer depends on somebody's recollection, you have a process, not a control.

Ready to go further? Download the five-part legal series, or send us a workflow and we will tell you plainly whether it is worth building.