A walk through Anthropic's new Legal Plugin

A walk through Anthropic’s new Legal Plugin

On 12 May, Anthropic released Claude for Legal. Twelve practice-area plugins, nineteen connectors, a deployment guide aimed at general counsel, and a public webinar. The IP plugin alone has twelve skills, one scheduled agent, five preconfigured MCP servers. What many do not realise: this is not a standalone legal-tech application but, at its core, a prompt library. And it is publicly available on GitHub under Apache 2.0. (repository link). There is no reason not to take a closer look and learn from it for your own prompt engineering.

I have spent several days reading the IP plugin file by file. The full write-up, with a table of all twelve skills and the eight design patterns I extracted, is in the new Wissens-Artikel on this site. What I want to do here is take you on a short walk through the parts that stuck with me.

The FTO skill opens with a warning

The first thing the FTO-triage skill does is put up a warning in capitals and bold:

This is not a freedom-to-operate opinion. A formal FTO opinion requires a comprehensive search, full claim construction, and element-by-element infringement analysis by registered patent counsel. Patent infringement is strict liability; willful infringement triples damages. A “no obvious blocking patents” result means the triage didn’t find one — it does not mean the product is clear. — ip-legal/skills/fto-triage/SKILL.md

A reminder of what language models can do (error-prone first pass) and what a patent attorney does (opinion).

What follows is equally instructive. The skill’s calibration rule:

Under-flagging a blocking patent is a one-way door — a product launched, a deposition a year later, treble damages on the table. Over-flagging is a two-way door — the attorney narrows the list in a read-through. Stay on the two-way door side. Always. — ip-legal/skills/fto-triage/SKILL.md

A rule any patent practitioner can confirm the relevance of. If only there were no hallucinations…

A disclaimer for EU use

Important for use outside the US: all twelve plugins carry the same jurisdictional guardrail in their CLAUDE.md:

“Applying US doctrine here would give you a wrong answer that looks right.”

What we can learn for our own prompting

The most interesting parts of the plugin are not the disclaimers but the rules against typical AI weaknesses. Three of them transfer directly into your own prompt practice:

1. Label sources, don’t invent them. The plugin requires every statement to carry an origin tag. The core rule:

“Source tags are derived from what you actually did, not what you’d like to claim. [model knowledge — verify] — everything else. This is the default. If you didn’t retrieve it, it’s model knowledge, no matter how confident you are.” — ip-legal/CLAUDE.md

2. Name doubt instead of staying silent. When the model knows a citation but has doubts about its validity, it must not simply stay quiet:

“Silence about known doubt is as misleading as confident assertion.” — ip-legal/CLAUDE.md

3. Verify the user’s premises before building on them. Language models tend to accept every user statement as correct. The plugin demands the opposite:

“Verify user-stated legal facts before building on them. A wrong premise propagated through three paragraphs of analysis is harder to catch than a wrong premise flagged at sentence one.” — ip-legal/CLAUDE.md

The redline principle

One more stop on the walk. The ip-clause-review skill on redline granularity:

Edit at the smallest possible granularity. A redline is a negotiation artifact, not a rewrite. Wholesale clause replacement signals “we threw out your drafting.” Surgical redlines — strike a word, insert a phrase, restructure a subclause — signal “we have specific asks” and are faster to read, understand, and accept. — ip-legal/skills/ip-clause-review/SKILL.md

AI-generated contract language tends to rewrite rather than edit. The plugin pushes against that.

What this means for your practice

Two things. First, the architecture is worth studying even if you never install the plugin. The source-tag taxonomy, the three-valued no-silent-supplement rule, the severity floor, the decision-tree close. These patterns are jurisdiction-agnostic. They address the same problems we regularly discuss in this newsletter. They are already folded into the prompt-tools on this site (FTO, opposition, scope-of-protection, office-action response all carry these disciplines now). The Wissens-Artikel has the full list with verbatim quotes you can expand.

Second, the gap the plugin leaves open is the EPC-side answer. Patent claim drafting is explicitly excluded: “Not claim drafting — this plugin doesn’t go there.” (ip-legal/skills/cold-start-interview/SKILL.md) An application where language models, in practice, do not produce adequate results.

On confidentiality

Claude for Legal runs on Anthropic’s cloud. A local install of the plugin file changes nothing about that; the model still sees every input. For real client data, the choice is either a cloud instance that meets the applicable professional-confidentiality requirements, or running the plugin against a locally hosted model (which Claude Code supports, with the usual performance trade-offs).^*

Read the source files and take the architecture as a blueprint for your own prompt library.

Resources

1. The Anthropic IP Plugin · Wissens-Artikel (DE/EN)
2. claude-for-legal on GitHub (Apache 2.0)
3. LTO: Das ist keine ‘Revolution’

^* Processing of confidential content is subject to professional-confidentiality requirements (BRAK, epi guidelines, § 203 StGB and their equivalents). These can be met by a locally hosted model or, in individual cases, by a trustworthy or self-hosted cloud instance. This does not constitute legal advice.

Enjoy exploring!

Sebastian Goebel, Founding Partner, Bösherz Goebel