AI-native law firms explained: Crosby, Garfield.law, Eudia and what the model means for you

At 7pm on a closing night in early 2026, a 40-page counterparty redline landed in a two-person firm’s inbox. By 11pm the lawyer had counter-language for every material change, tracked changes in the Word file attributed to his own name, and a client email explaining it. Total time, he wrote, “under an hour, of which about 30 minutes is my own thinking”. Zack Shapiro’s account of that evening, posted on X on 27 February 2026, was read by more than 7 million people. It is the most-read description of an AI-native law firm there is, and it is where the phrase “Claude-native” comes from.

A month later Crosby, a venture-backed firm where lawyers and engineers sit at alternating desks, raised a $60M Series B on a 58-minute median contract turnaround and no billable hour at all. Both get called an AI-native law firm. They are not the same thing; the difference is the useful part.

What “AI-native” means: the remove-the-AI test

The cleanest definition comes from the AI-Native Law Firm report: “in a bolted-on firm, removing the AI tools is an inconvenience; in an AI-native firm, it breaks the operation”.

Two cautions. About 90% of legal dollars still flow through hourly arrangements; Crosby’s co-founder Ryan Daniel concedes the billable hour is “just really durable”. And the report’s non-negotiables include one most firms skip: “lawyers are trained to verify rather than merely edit”.

The post and why it spread

Shapiro’s claim: a general-purpose model, configured well, beats vertical legal AI for a small firm. Artificial Lawyer’s read: “possibly 7 million people didn’t know you could do the above things with Claude”.

Three modes: Chat, Cowork, Code

Chat everyone knows. Cowork is the one Shapiro says most lawyers have not tried: “I point Claude at a folder on my computer, give it a task, and it goes and does it.” Code is the terminal, where he built a tool that reads contracts aloud for his commute.

Ernie Svenson moved from ChatGPT for the same reason, file access: “I hadn’t switched because ChatGPT failed me. I switched because I finally saw what I’d been missing.” He had Claude interview him with 40 questions about how he works; the answers became “a master skill that now sits in my top-level prompt”. The agent explainer explains why Cowork is closer to an agent than a chatbot.

Skills: encoding judgement once

Shapiro’s sharpest line: “The difference between a firm playbook and an individual lawyer’s encoded judgment is the difference between giving someone a recipe and teaching them how to cook.” A skill is a standing instruction file for one recurring task. His six run from contract review (severity ratings, a missing-provisions checklist) to research, whose mandatory self-review checks that “every cited authority actually says what the memo claims”. He built them by uploading Anthropic’s skills guide and asking, “based on the hundreds of conversations we’ve had together… what are the skills that would have the greatest impact on my practice?” Anthropic’s open-source Claude for Legal repository now ships 93 named agents across 12 practice-area plugins; its README warns that “Skipping setup is the single most common reason a skill produces generic output”. One caveat for European readers: the plugins are calibrated for common law.

I want to build a reusable instruction file (a "skill") for [contract review from the customer side] in my [practice area] practice in [jurisdiction]. Do not write it yet. Interview me in batches of ten questions, up to forty, about: my clients and their commercial posture; the deviations I always flag and the ones I concede; my house style; how to handle citations (tag every authority [VERIFY]); what must never appear in an output; how to behave when unsure. Then draft the skill, safety rules first, under 600 words.

The bad prompt versus the good prompt

The post’s own before-and-after. The bad prompt is three words: “review this contract”. The good one:

“review this services agreement from the vendor’s perspective. Flag provisions where the customer shifted risk beyond market norms for this type of deal. Check for missing provisions that should be present, including limitation of liability, IP ownership, data handling, and termination for convenience. Produce a severity-rated summary with specific counter-language for each high-severity issue. Note that the vendor has limited negotiating leverage and wants to close the deal, so recommendations should focus on provisions worth fighting for versus provisions to concede gracefully.” — Zack Shapiro, Rains LLP, 27 February 2026

Side, standard, checklist, output format, commercial posture: everything a partner would tell an associate. In Shapiro’s words, “The entire gap between ‘AI is a toy’ and ‘AI changed my practice’ lives in the quality of your instructions.” Generalised, with the verification line the original lacks:

Review the attached [agreement type] from the perspective of [our client: vendor/customer], a [size, sector] business that [has little bargaining power and wants to close within two weeks]. Apply [jurisdiction] law. Flag every provision where [the counterparty] shifts risk beyond a typical mid-market position, quoting the operative words. Check for missing provisions: [limitation of liability, IP ownership, data handling, termination for convenience]. Produce a severity-rated table with counter-language for each high-severity issue, separating "worth fighting for" from "concede gracefully". Cite no case or statute unless certain it exists; otherwise write "unverified".

The Projects and custom GPTs guide makes it persistent; the prompt library has the rest.

“Not using these tools is the harder position to defend”

That is Shapiro’s professional-responsibility claim, and it has support: the UK Jurisdiction Taskforce’s July 2026 legal statement says at paragraph 67 that “a professional could also be liable for failing to use AI for a task when a professional exercising reasonable care and skill would have done so”. His line for senior lawyers: “If you’ve spent 10 or 20 years developing judgment in your practice area, you are sitting on exactly the asset that AI makes more valuable, not less.”

It is not a licence to stop reading. Lawyers beat every tool on redlining in Vals’ February 2025 benchmark (79.7% against Harvey’s 65.0%); the replacement debate turns on that residue of judgement.

The rebuttal: pots versus rice cookers

Noah Waisberg, who founded Kira and now runs Zuva, answered the post in a piece titled “Is Legal AI Cooked?”: “Saying ‘foundation models mean legal AI is cooked’ is like saying ‘rice cookers are totally useless when you have a pot.’” His better argument is about what you are building: “as you invest more and more into your Claude prompts and workflows, what you’re building starts to look a lot like software… except without the QA, the versioning, the user feedback loops, or the ability to survive someone leaving the firm.”

The market is moving Shapiro’s way while Waisberg’s warning stands: Clio’s 2025 report found only 40% of firms using a legal-specific tool, down from 58% in 2024. The general versus legal-specific comparison sets out when the rice cooker is worth its price.

Confidentiality: Team plan, no consumer accounts

Everything above assumes a commercial plan. Since 28 August 2025, Claude Free, Pro and Max use conversations for training by default, with five-year retention for those who opt in. Claude Team, Enterprise and the API do not train on inputs by default; admins should also disable “Rate chats”, because thumbs-up feedback stores the conversation for five years. For EU firms: Claude’s first-party workspaces are stored in the US; EU processing runs only through AWS Bedrock or Google Vertex.

What a five-lawyer firm can replicate in 30 days

No coding, Team plan assumed; the 30-day plan for small firms is the general version.

Week 1: settings and paper. Team seats for everyone; training off; “Rate chats” off; an engagement-letter clause covering the tools, no-training terms, human review and billing (actual time only).

Week 2: the interview. Each lawyer runs the cold-start interview above for their most repeated task and saves the result as a skill or Project instruction; build one anonymised clause bank for the firm’s top contract type.

Week 3: one workflow, tested on history. Run the contract-review skill on three closed matters where you know what the review found, and compare; Harvey’s pilot advice applies to home-built tools too: test on “historic matters… where you already know the outcome”.

Week 4: Cowork on a real folder. Point Cowork at an anonymised matter folder, run the redline-to-response workflow, and check the tracked changes against Word’s own compare before anything leaves the building.

Compare <our_draft> with <their_markup>. Table every change, including deletions and moved text: Clause | Our language | Their language | Effect on [our client] in one sentence | Severity (Dealbreaker / Significant / Minor / Cosmetic) | Recommended response (Accept / Counter with: [text] / Reject with reason). Then identify tensions their changes create with provisions they did not touch, and draft a one-paragraph, neutral cover note to their counsel. Do not apply any edit to the document until I confirm each row.

What not to copy

Unread tracked changes. Lawyers still beat the tools at redlining; the changes are a draft attributed to you.

Persona inflation. A March 2026 thread of “You are a senior corporate attorney at Skadden…” prompts drew a backlash; Artificial Lawyer asked whether Claude “simply read ‘Wachtell’ and thought ‘OK, this means do the contract in the style of any large commercial law firm’”. Shapiro’s prompt states side and posture, not a costume.

Home-built tools holding client data. “Lawyers should not trust apps they vibe-code with their clients’ confidential information unless they have the technical expertise to deploy their tools securely” (Version Story); the vibe-coding guide covers the safe path.

Dropping the verification skill. The research skill’s mandatory check of every authority is the part that answers the sanctions cases; keep it whatever else you leave out.

Where to go next: the careers and future hub has the wider picture; the Claude for lawyers guide goes deeper on Projects, Cowork and Claude for Word. Seven million readers “didn’t know you could do the above things with Claude”; AI Lab for Lawyers shows you, with Cowork open on screen and the Team-plan guardrails set first.