Sign in Subscribe

Claude for Legal Is Huge!

Claude for Legal Is Huge!
Claude for Legal is a massively disruptive piece of AI infrastructure
Anthropic just rolled out Claude for Legal, and even if you're buried taking, grading or just having recurrent nightmares about exams, this post deserves ten minutes. At least three things matter for the academy: a law-student plugin with a summa cum laude portfolio of skills; a free CourtListener connector that does traditional legal research and more while quietly routing around the Westlaw/Lexis duopoly; and a legal-clinic plugin that turns supervisory pedagogy into configuration and could bend the cost curve of clinical and legal-aid work. There is much more and none of it is finished — but the trajectory is unmistakable. Access-to-justice, clinical economics, and how and to whom we teach law itself are all about to look different faster than we can imagine. This is only the beginning.

It was inevitable. Claude has entered the legal tech market big time. It's a huge development. In this blog post I am just going to alert readers to what has happened and show a few examples. But probing all the capabilities of Claude for Legal is going to span a number of posts that I'll try to get out over the next week or two. (Gotta grade those exams 😦)

The Law Student Plugin

Claude has added a "law student plugin" that contains a portfolio of skills that law students should find incredibly useful. Maybe professors too! Here's a listing.

To install the skills, you have to religiously follow the instructions in the video contained at this website. https://github.com/anthropics/claude-for-legal/blob/main/QUICKSTART.md Watch it carefully! (I messed up the first few times due to impatience and incorrect assumptions about how the UI works).

I'll let the new skill create a Socratic discussion of separation of powers.

Let's see how its legal-writing skill performs. I had Grok draft a deliberately cruddy answer to a hypothetical exam question involving separation of powers. Here's a screen capture of the beginning of Claude's response.

Let's see if it can forecast an exam question I might ask? I gave it my most recent constitutional law exam and wrote:


/exam-forecast Forecast a multiple choice question this fiendish professor might write.

After having conducted this accurate analysis of the fiendish professor's questioning style, it comes up with a future question.

Forecast question

A Chandler-style MCQ he very plausibly writes for next year — weaponizing Trump v. Anderson (2024), which is conspicuously absent from this year's bank despite being the Court's marquee structural decision of that term:

Question (forecast)

Following allegations of election-related misconduct, the Secretary of State of Franklin, after a contested administrative hearing, determines that former Senator Marcus Holloway "engaged in insurrection or rebellion against the United States" within the meaning of Section 3 of the Fourteenth Amendment. Franklin law provides that the Secretary "shall exclude from the ballot any candidate she determines to be constitutionally ineligible to hold the office sought." Acting under that authority, the Secretary removes Holloway from Franklin's presidential primary ballot. No federal statute purports to authorize state enforcement of Section 3 against federal candidates, though 18 U.S.C. § 2383 makes it a federal crime to engage in insurrection or rebellion. Holloway sues in federal court, arguing Franklin lacks constitutional authority to disqualify him. How should the court rule?

(a) For Franklin. Article II's Electors Clause grants States plenary authority to determine the "manner" of appointing presidential electors, which necessarily includes the authority to determine which candidates may appear on the State's ballot. Section 3 imposes a textual qualification, and the State's authority to enforce a textual qualification follows from its authority to administer the election. See Chiafalo v. Washington.

(b) For Holloway. Section 3 of the Fourteenth Amendment may not be enforced against candidates for federal office absent legislation enacted by Congress under Section 5. The Constitution channels enforcement of Section 3 against federal candidates through Congress rather than the States in order to avoid a patchwork of inconsistent state-by-state disqualifications that would fracture the uniform national character of federal elections. Trump v. Anderson, 601 U.S. 100 (2024).

(c) For Holloway. Whether a particular individual "engaged in insurrection" within the meaning of Section 3 is a nonjusticiable political question committed to Congress by Article I, § 5, which makes each House the Judge of "the Elections, Returns and Qualifications of its own Members" — and the matter is therefore outside the competence of any State or court.

(d) For Franklin, but only because 18 U.S.C. § 2383 supplies the congressional authorization required to enforce Section 3 against federal candidates; absent that statute, Franklin would lack power to disqualify Holloway, but the statute cures the defect.

Correct answer: (b).

There's tons more to explore here in the law student plugin. Claude for Legal is a giant release. But let's move to another capability released today.

The CourtListener Connector

One huge feature for legal educators, especially those without large budgets, is the ability to connect to CourtListener for free. Other products have claimed before to offer this kind of connection to this vast legal database, but my experience with earlier versions was not stellar. Here, you can see me using the new connector to search for Texas cases on the duty to settle decided after 2020. I am not using my paid Midpage subscription. I am using the free CourtListener MCP connector.

That matters. Students may no longer need to spend $25 a month for decent access to grounded AI. The free CourtListener version may not be quite as powerful, but it can still reduce hallucination risk substantially and produce grounded answers within the familiar framework of a consumer large language model that already offers dozens of other capabilities.

The CourtListener skill can retrieve cases that it has found and discuss them.

So it can apparently find cases, retrieve them, and comment upon them. But can it check if the cases it discovers are still good law? Yes! Here is my deliberately cruddy prompt:

There is a Texas insurance case called something like Farmers v. Soriano from the 1990s. Can you find the case and see if it has any negative treatment.

I located the case and checked for negative treatment. There are actually two opinions in this litigation:

The Texas Supreme Court decision — Texas Farmers Insurance Co. v. Soriano, 881 S.W.2d 312 (Tex. 1994) — is the one you're almost certainly thinking of. It reversed the San Antonio Court of Appeals (844 S.W.2d 808) and established what is now usually called the "Soriano rule": in a multiple-claimant situation, an insurer that in good faith settles with one or more claimants in a way that exhausts (or substantially exhausts) policy proceeds is not liable in Stowers to the insured even if remaining claimants are left without recovery. CourtListener shows 106 citing opinions.

Negative treatment in CourtListener's data: None of the standard markers turn up. The case has not been overruled, abrogated, or disapproved. I ran searches for "Soriano" paired with "overruled," "abrogated," "declined to follow," "distinguishable," "rejected," "criticized," and "disagree." The hits that came back are overwhelmingly (a) cases involving unrelated parties named Soriano (immigration, criminal, workers'-comp), or (b) opinions citing Soriano favorably. A spot-check confirms this: American Guarantee & Liability v. ACE American Insurance, 990 F.3d 842, 848–49 & n.16 (5th Cir. 2021), recites the Soriano three-prong test as canonical Texas Stowers law; Westport Insurance v. Pennsylvania National Mutual, 117 F.4th 653 (5th Cir. 2024), and OneBeacon v. T. Wade Welch & Associates likewise apply it as good law.

The only "negative" voice inside the Soriano file itself is the original concurrence by Justice Hightower (joined by Doggett, Gammage, and Spector in part) stating "I disagree with the court's analysis of an insurer's duty to settle (including the requirement of a formal settlement demand within policy limits)," cross-referencing his dissent in American Physicians Insurance Exchange v. Garcia, 876 S.W.2d 842 (Tex. 1994). That is intra-court disagreement at the time of decision, not subsequent negative treatment, and it never commanded a majority.

Two caveats. First, CourtListener does not maintain a Shepard's- or KeyCite-style negative-treatment flag system — the absence of negative-treatment hits in my keyword searches means I found nothing, not that nothing exists. Second, my searches showed several later cases that distinguish Soriano on its facts (e.g., refusing to extend its safe harbor where the insurer's settlement did not actually exhaust limits, or where the offer was not a proper Stowers demand). Distinguishing is not negative treatment in the Shepard's sense, but it is worth being aware of when relying on the case. For a definitive answer in a brief, I'd still run KeyCite or Shepard's.

Sources:

Students and other legal professionals can use the CourtListener connector to summarize a body of law. Here's my prompt. Again, I used absolutely minimal effort.

Retrieve all the post 2020 cases that you found and create a 300 word essay on how they change or clarify the duty to settle in Texas. I will permit you to use the web to research how the baseline Stowers doctrine works in texas. Put your answer in an html div with a light blue background.

A few minutes later, I get this:

Let me pull the Westport opinion and do a quick web check on baseline Stowers before drafting.

The Stowers Doctrine After 2020: Clarification, Not Revolution

The Stowers doctrine, rooted in G.A. Stowers Furniture Co. v. American Indemnity Co., 15 S.W.2d 544 (Tex. Comm'n App. 1929), requires a liability insurer to exercise ordinary care in responding to a third-party settlement demand or face liability for any excess judgment. American Physicians Insurance Exchange v. Garcia and Texas Farmers Insurance Co. v. Soriano (both 1994) fix three triggering prerequisites: the claim is within coverage; the demand is within policy limits; and an ordinarily prudent insurer would accept it given exposure. The demand must be unconditional, for a sum certain, and propose a full release.

The post-2020 cases sharpen these edges without rewriting the framework. American Guarantee & Liability v. ACE American Insurance, 990 F.3d 842 (5th Cir. 2021), reaffirms that ambiguity defeats Stowers — there, an offer of "$1.9MM to $2.0MM with costs" failed because the demand did not clearly state a sum certain. The same opinion Erie-guesses that a lump-sum offer made by a parent as next friend for minor children is not inherently conditional merely because post-acceptance judicial approval will be required, modestly expanding the universe of triggering offers. And it holds that arguments about appellate-reversal prospects must be preserved in the trial court or are waived.

Westport Insurance v. Pennsylvania National Mutual, 117 F.4th 653 (5th Cir. 2024), clarifies the primary-versus-excess interaction: an excess insurer's later contractual breach does not excuse the primary's earlier Stowers breach, and Stowers is not a freestanding defense to an excess carrier's duty to indemnify. In re State Farm Mutual Automobile Insurance Co. (Tex. App.—Fort Worth Nov. 19, 2020) is largely dicta but reminds readers that Murray abandoned Arnold's Stowers analogy as the doctrinal bridge between first-party UM/UIM and third-party bad faith. The post-2020 picture is incremental tightening, not doctrinal upheaval.

It can also create precedent graphs (although I am not sure I would vouch for this one).

Use the CourtListener MCP to Make a network visualization that connects the citations between Myers v. United States and Trump v. Slaughter in the case law database."

But CourtListener goes beyond judicial opinions. It also permits access to docket entries. Here's an example of a query and a screen capture of part of the response.

use courtlistener to find docket entries from 2026 in the Nathan v. Alamo Heights case in the 5th Circuit. Put the answer in a markdown file.

I had often wondered if Claude or one of the other frontier AI labs would use part of their billions to crash through the eroding Westlaw/Lexis duopoly and simply buy a legal database for their users. I had thought (and continue to think) it would be a great idea. By now connecting to a pre-existing database, however, and developing an intelligent connector, Claude may have gotten at least halfway there for free. It's the first frontier AI lab out of the gate in this effort but I suspect not the last.

Another important part of today's release is the legal-clinic plugin. This one isn't aimed at the solo or the BigLaw associate — it's aimed squarely at the law school clinic supervisor and her students, and (I think) at the legal aid lawyer trying to do five times as much work as her funding allows. Sixteen skills in all. Rather than march through every one, let me pull out the three I think do the most work, and just list the rest.

cold-start-interview. The professor's one-time clinic setup, and the foundation everything else rests on. Captures practice areas, jurisdiction, supervision style (formal review queue / configurable flags / lighter-touch), and ingests the handbook, filing guides, local court rules, intake forms, and a scrubbed example file. Writes a CLAUDE.md every other skill reads. This is what lets the professor encode her pedagogy once and have it scale across every student-facing skill without her having to repeat it. Quietly the most important skill in the bundle.

draft. First drafts of common clinic documents — asylum applications, eviction answers, protective-order petitions, demand letters — with jurisdiction-aware formatting and [FACT NEEDED] / [VERIFY] / [UNCERTAIN] flags planted inline where the work isn't done yet. Loud AI-assisted label at the top, automatic supervision routing, no pretense that this is anything but a starting point. This is the workhorse. It will be the skill students use most and the one that compresses the lowest-value hours of clinic work into minutes.

semester-handoff. The mirror of /ramp. At the end of the term it reads deadlines.yaml and the per-case comms logs and generates a transition memo for every active case plus a cohort summary. The institutional knowledge that used to walk out the door with the graduating 2L — what was promised the client last March, where the research dead-ended, which deadline is sitting on July 14 — just stays put. The most distinctive structural contribution in the plugin and the one I'd build a clinic around.

The rest, briefly: build-guide (lets the supervisor configure pedagogy per practice area), client-comms-log (append-only per-case comms record), client-intake (structured intake with conflict and issue-spotting), client-letter (routine correspondence from templates), customize (change one thing without re-running cold-start), deadlines (the cross-case deadline tracker), memo (IRAC scaffold with blank Application and Conclusion — pedagogically smart), ramp (student onboarding from the handbook the professor uploaded), research-start ("leads not authorities" research roadmap), status (case status reshaped for client / internal / court audiences), supervisor-review-queue (the professor's review inbox when she's chosen formal-queue supervision). Two more — form-generation and plain-language-letters — are deprecated, folded into /draft and /client-letter+/status respectively.

Why this could matter a lot

Now let me put my professor hat on and tell you why I think this release is more important than the law student plugin everybody is going to focus on.

Law school clinics — and the legal aid organizations that share their economics — are caught in an unforgiving bind. Every case needs a credentialed attorney to supervise an inexperienced student. That supervision is the largest cost driver in clinical legal education and the binding constraint on how much pro bono work a legal aid office can actually deliver. Students often churn in semester cycles. Knowledge walks out the door every December and every May. Intake produces administrative drag. Routine correspondence absorbs hours that should be going to advocacy. The result is a chronic gap between the volume of unmet legal need and the throughput of the institutions trying to meet it. This plugin attacks every link in that chain.

The first thing it does is turn the supervising attorney's pedagogy into configuration. cold-start-interview and build-guide let her encode her practice areas, jurisdiction, supervision posture, and teaching philosophy once. Every downstream skill — draft, memo, client-intake, status — then reads from that same CLAUDE.md. supervisor-review-queue enforces the review gates she chose. Her judgment scales without her having to repeat it.

The second thing it does is compress the structurally similar, lowest-leverage hours of clinic work. client-intake produces a formatted summary in minutes instead of half an afternoon. draft generates a jurisdiction-aware first draft of an eviction answer or a protective-order petition with the missing facts visibly flagged. research-start hands the 2L a roadmap instead of letting her flail at the Westlaw bar for two hours. client-letter and status automate the dozens of routine touches that used to be evening work. None of these outputs is a final product — every one is loudly labeled AI-assisted and routes back through supervision — but the time formerly consumed by the lowest-value tasks now belongs to the irreducibly human ones: client counseling, strategy, advocacy.

The third thing — and the one I find most quietly impressive — is what deadlines, client-comms-log, and semester-handoff do together. They build a persistent operational substrate that is itself the answer to the semester-churn problem. The 2L who arrives in August inherits per-case handoff memos and a cohort summary built automatically from the deadline file and the comms logs. ramp then teaches her the clinic she is actually entering, from the very handbook her professor configured at cold-start. The institutional memory that used to evaporate with each graduating cohort just doesn't anymore.

Here's an example of what the deadline skill generates.

/legal-clinic:deadlines Use the skill to alert me to the series of deadlines my client is likely to face in a Houston eviction case. She got served with a notice to vacate yesterday but she believes she has proof (a) that she is current in her rent and (b) that the landlord threatened to report her to ICE if she did not pay in full. That might be extortion.

This is a useful answer but it looks as if the skill could be improved. Fortunately skill files can generally be forked so that a Texas clinic could likely adapt it for its own use by using the Illinois or California material as a template.

By the way, to test out whether the skill files are as portable as advertised, I ran it in OpenAI's Codex rather than Claude. It worked a little differently, but it did function. This fact leads me to believe that many of the skills now available for Claude for Legal may also work at least partly inside other large language models.

Here's a skill file. Use it to alert me to the series of deadlines my client is likely to face in a Houston eviction case. She got served with a notice to vacate yesterday but she believes she has proof (a) that she is current in her rent and (b) that the landlord threatened to report her to ICE if she did not pay in full. That might be extortion.I used the deadline-tracking approach from your skill file and set a single thread alert to run daily from May 14, 2026 through May 26, 2026. It will prompt you to watch for a filing, citation, service date, and trial date so the later deadlines can be calendared from real trigger dates rather than guesses.

I used the deadline-tracking approach from your skill file and set a single thread alert to run daily from May 14, 2026 through May 26, 2026. It will prompt you to watch for a filing, citation, service date, and trial date so the later deadlines can be calendared from real trigger dates rather than guesses.

For this Houston eviction matter, the immediate timeline is:

  1. May 11, 2026: notice to vacate served.
  2. May 14, 2026: likely first day the landlord could file, assuming this is the default three-day Texas notice period and the lease does not require a longer period. Texas now counts the notice period from the day of delivery. This is an inference from the statute, not a court-confirmed case date. ([jp.hctx.net](https://www.jp.hctx.net/evictions/faq.htm))
  3. If the landlord files:
  4. If the landlord seeks summary disposition under the 2026 rules and a response deadline applies from the papers served, that response can be very short, including a 4-day response period in the rule text quoted by the Texas Supreme Court materials. The exact applicability should be checked against the filed motion and case type.
  5. If judgment is entered against the tenant:
    • Appeal is due within 5 days after judgment is signed.
    • A writ of possession generally cannot issue before the 6th day after judgment or before the appeal deadline passes, whichever is later.

Two merits points are worth flagging now:

  • Because the Texas statute changed effective January 1, 2026, if the eviction is based solely on nonpayment and she was not late or delinquent before the month of the notice, the notice must be in the form of a notice to pay rent or vacate, not merely a bare notice to vacate.
  • Her proof that rent is current is central. The ICE threat should also be preserved exactly as received, but I have not classified it as extortion here without a separate legal analysis.

The next useful trigger is any filed petition, citation, service date, or hearing notice. Once one of those exists, I can turn the rest into exact dates instead of conditional ones.

Stack these various skills and the cost curve of running a clinic bends. A clinic that supervised eight students on twenty cases could plausibly supervise eight students on thirty-five — or twelve students on the same twenty, with deeper case work per student. Legal aid offices outside law schools, operating under the same supervision constraint with even less slack, benefit in a smilar way. Housing, immigration, bankruptcy, family law — the document-heavy practice areas where the unmet-need gap is widest — are exactly where this plugin might hit hardest. But it might also enable new clinics in diverse fields: imagine the ERISA clinic, the maritime clinic, and other areas in which practical experience might benefit students and communities alike.

And more ...

There is more in the Claude for Legal announcement that is of great relevance to legal educators and the legal profession generally. One notable item is a partnership with Courtroom5, a platform that describes its mission as helping self-represented litigants. I have not yet been able to get its Claude connector to work, so any assessment has to remain provisional. But certainly the idea is important. If Claude can connect effectively with software designed for self-represented litigants, that could become a major access-to-justice development. AI may enable an averagely educated person to function much more like a competent legal advocate, at least in smaller matters where the alternative is often paying $400 or more per hour for a licensed attorney—or going without help altogether. We will explore the implications in future posts. For now, the central point is simple: a substantial increase in AI-assisted pro se litigation has already begun, and it is likely to matter a great deal for both the judiciary and legal education.

Future Posts

As I said at the beginning, Claude for Legal is a huge development. I want to write more. Unfortunately, I've got a stack of exams that do not seem to be grading themselves. And although it is super tempting just to throw them into AI and see what results, I am still old school enough this year to feel some obligation to grade them myself. But we're going to come back to many of the new features of Claude for Legal and a related development that I'm hoping I can access and test for myself, Claude for Microsoft 365, which lets you use Claude inside the apps with which most legal academics and legal professionals are most comfortable.

We often say, somewhat casually, that AI is advancing exponentially. But those can remain abstract words until you see what that means in practice: tools like Claude for Legal, projects like Suzie Law, Irys, or Legora aOS, and the seemingly endless stream of new legaltech products arriving almost daily. Then the pace of change stops feeling theoretical and starts hitting you in the face.

Claude for Legal is one of the clearest examples yet. I have only scratched the surface here, and I expect to return to it soon. But even this early look suggests that it is not merely another useful legal AI product. In my world of legal education and beyond, it is a deeply disruptive one.