Claude Skills for Legal Work: Build the Playbook Once

The Short Answer

A Claude Skill is a reusable package of instructions, and optionally templates, reference files, and scripts, that Claude loads on its own when a task matches the skill's description. Anthropic's Help Center defines Skills as extending Claude's capabilities "by giving it access to specialized knowledge and workflows," and they are available on every plan including Free (code execution must be enabled, and on Enterprise an Owner must have Skills enabled in organization settings) on claude.ai, and also carry over to the Microsoft 365 add-ins, Cowork, and Claude Code on plans that include those surfaces (Cowork and Claude Code require a paid plan), as of October 2026. For a law firm, a Skill is where the playbook goes: your deposition-summary format, your NDA fallback positions, your engagement-letter checklist, written once, versioned, shared with the whole firm, and applied consistently without anyone re-pasting a 400-word prompt. The rest of this post covers what a Skill is in concrete terms, how a firm builds one, how Anthropic's own legal toolkit is structured as Skills you can copy, and a worked example for one recurring task.

What a Skill Is, Concretely

Per Anthropic's custom-skills guide, a Skill is a folder containing a required SKILL.md file plus any supporting files. The SKILL.md opens with two required fields: a name (64 characters max; the Help Center calls it a human-friendly name and its example is in title case, but the Agent Skills spec requires lowercase letters, numbers, and hyphens matching the folder name, so that form is the safest) and a description (200 characters max per the Help Center, although the spec allows up to 1,024) that tells Claude what the skill does and when to use it. Below that, the body is ordinary markdown: the instructions Claude follows when the skill runs. Supporting files can include reference documents, templates, images, and, on surfaces with code execution, scripts. You zip the folder and upload it under Customize, then Skills.

Two properties distinguish a Skill from a long prompt or a Project instruction. First, it loads only when relevant: the description is what Claude reads to decide whether to invoke it, so a well-written description ("Use when summarizing a deposition transcript for a litigation file") is the most important line in the file. Second, it is portable: on Team and Enterprise plans the Help Center describes skills organized into Personal, Shared with you, and Organization sections, so a partner can build one and share it with colleagues or publish it to the firm's organization library, and an Owner can provision it to everyone.

A note on Styles, since many lawyers set one up in 2025: reporting from How-To Geek and others says Anthropic has retired Styles in favor of Skills, and as of October 2, 2026, Anthropic's personalization overview lists Instructions, Project instructions, and Skills as the three personalization features, with no Styles entry. The practical difference: a Style applied to every message; a Skill is invoked when Claude judges it relevant. If you relied on a Style for house voice, rebuild it as a Skill with a description broad enough to trigger on drafting tasks, and put the voice rules in account-level Instructions as a backstop.

Skills in Claude Code and Cowork

The same folder format works in Claude Code, with more options. Claude Code's skills documentation describes skills living in a personal folder, a project folder (committed alongside the files so the whole team gets it), inside a plugin, or in enterprise-managed and other locations; it supports slash-command invocation (/deposition-summary) as well as automatic invocation, and extra frontmatter fields such as one that restricts a skill to user-only invocation. The docs note that claude.ai-uploaded skills support only the smaller set of frontmatter fields in the Agent Skills spec (name, description, license, compatibility, metadata, allowed-tools), and that a skill enabled for your claude.ai account becomes available in Cowork and cloud sessions. For a non-technical firm, the practical reading is: build skills as plain markdown plus templates, upload them to claude.ai, and they will work everywhere you use Claude, including Cowork where the legal plugins run. See Claude Cowork for law firms for that surface, and Claude Code for lawyers for when the extra fields matter.

How a Firm Builds Its Own

A legal Skill has three parts, and firms that skip the first one get generic output.

1. The playbook. Your actual positions and standards, written as rules. For an NDA skill: which clauses you accept as-is, which you always push back on, your fallback for each, and the two or three deal-breakers. For a deposition-summary skill: the format (page:line cites, topic headings, a "notable admissions" section), what to flag, what to leave out. This is the part only your firm can write, and it is the whole value.
2. The templates. The output shape, as a file in the skill folder: the summary template, the client-letter shell, the redline memo format. Claude fills templates far more reliably than it invents structure.
3. The instructions. The SKILL.md body: the steps in order, what to do when information is missing (ask, do not assume), the verification reminders ("mark every citation UNVERIFIED until checked against the source"), and the tone rules. Keep it short; the docs' guidance is to start with simple markdown before adding scripts, and to keep each skill to a single workflow.

Then the description. Write it as a trigger condition, not a summary: "Use when asked to review, redline, or triage an NDA or confidentiality agreement" will fire correctly; "NDA helper" will not. Test with three or four real requests and check that the skill activates. If it does not, sharpen the description before touching anything else.

A firm on Team should build in one partner's Personal skills, test for a week, then share to the Organization. Version with a dated changelog line at the top of SKILL.md (keep the name itself a stable lowercase, hyphenated identifier), because you will revise it, and you want to know which version produced a given draft.

The Official Legal Toolkit Is Skills You Can Copy

Anthropic's Claude for Legal repository is, structurally, twelve Anthropic-built practice-area plugins (plus a partner-maintained CoCounsel plugin), each bundling a set of skills (the repository describes skills as domain expertise and step-by-step methods Claude draws on automatically, plus slash actions you trigger explicitly, such as a vendor-agreement review or a DSAR response) and a template practice profile that every skill reads from; many of the plugins also include scheduled agents. Because the license is Apache 2.0, every one of those skills is a worked example you may read, copy, and rewrite. The most useful move for a small firm is not to install all twelve plugins; it is to open the litigation or commercial plugin, read how Anthropic structured a skill (the description, the step order, the way templates are referenced, the "flag as unverified" conventions), and build your own skills in the same shape for the practice areas the toolkit does not cover. None of the twelve is a dedicated plugin for family law, immigration, criminal defense, or personal injury practice as of October 2026 (the legal-clinic plugin touches some of those areas, but it is built for law school clinics); the structure transfers directly. Our setup guide covers installation; the plugins themselves are free under the Apache 2.0 license, but Cowork requires a paid Claude plan and some connectors (such as CoCounsel and Ironclad) require the vendor's own subscription; plan pricing is in Claude for Legal pricing.

Worked Example: A Deposition Summary Skill

Here is a complete, minimal skill for one recurring litigation task. Adapt the playbook section to your practice; the structure is the point.

Folder: deposition-summary/ containing SKILL.md and summary-template.md.

SKILL.md frontmatter:

• name: deposition-summary
• description: Use when asked to summarize, digest, or index a deposition transcript for a litigation matter. Produces a page:line-cited topical summary in the firm template.

SKILL.md body (abridged):

1. Read the entire transcript before writing anything. If the transcript is incomplete or pages are missing, say so and stop.
2. Produce the summary using summary-template.md. Every factual statement carries a page:line cite in the form 45:12-46:3.
3. Organize by topic, not chronology, unless the user asks for chronological. Suggested topics are in the template; add or remove as the testimony requires.
4. Include a "Notable Admissions" section and an "Inconsistencies With Prior Statements" section; leave the second empty with a note if no prior statements were provided rather than inferring.
5. Do not characterize credibility. Report what was said.
6. Do not summarize objections or colloquy except where an instruction not to answer was given; list those separately with cites.
7. Close with an "Open Items for Attorney Review" list: ambiguous testimony, exhibits referenced but not provided, and any cite you were unable to confirm.
8. Every summary is a draft for attorney review; state that in the header.

summary-template.md: a header block (case, deponent, date, taken by, summary prepared date, "DRAFT FOR ATTORNEY REVIEW"), topic sections, the two special sections, and the open-items list.

Zip the folder, upload under Customize, then Skills, enable it, and ask Claude to summarize a transcript. It should invoke the skill without being told. Compare the first output against the same transcript run through our unstructured deposition summary prompt: the difference is consistency across every summary the firm produces, plus the open-items discipline that catches what the model was unsure of. From there, the brief and transcript workflow shows how to chain the summary into a chronology.

Where Skills Fit With Projects, Instructions, and Prompts

Lawyers now have four places to put standing guidance, and the overlap causes confusion. Account-level Instructions apply everywhere and should hold only the universal rules (jurisdiction, voice, "flag unverified citations"). Project instructions hold matter- or practice-specific context for one workspace. Skills hold repeatable workflows that produce a specific output, and they are the only one of the four that carries templates and files. Prompts are what you type on the day. If a prompt keeps getting long and keeps getting reused, it is a Skill waiting to be written, a point we make from the other direction in why over-prompting hurts current models.