The Judgment Is Yours

Zack Shapiro, a transactional attorney running a two-person boutique firm, published a detailed account of how he's built his entire practice around Claude rather than any of the specialised legal AI products currently fighting for market share. It's long, specific, and worth reading in full — not because everything in it holds up, but because the parts that do hold up matter a lot.

The headline claim is that his two-person firm competes with firms a hundred times its size. Fine. Every small-shop operator says that. What makes this piece different is the detail. He describes uploading a 40-page redlined agreement and having Claude generate real tracked changes at the XML level inside the .docx — attributed to his name, formatted correctly, openable in Word by opposing counsel who'd never know it wasn't done by hand. He describes real-time cross-referencing of contract provisions during a live negotiation, catching contradictions between a counterparty's demand letter and their own previously drafted side letter. He describes a research workflow with a built-in self-verification step specifically designed to catch hallucinated citations — the problem that got other lawyers sanctioned and made national news.

This is not "I asked ChatGPT to summarise a contract and it was pretty good." This is a practitioner who has spent serious time building custom instruction architectures that encode his own analytical judgment, and who uses the tool's ability to write and execute code as a core part of his legal workflow. The distinction he draws between specialised legal AI products and a well-configured general-purpose model is the sharpest observation in the piece: the wrappers are selling you a template library dressed up as intelligence, when the actual value was never in the template. It was in what the lawyer did with it.

He's right about that. And the upstream implication — that any professional whose value comes from judgment rather than document production should be paying attention to this — is correct too.

But there's a confidence in the piece that outruns its evidence.

Shapiro frames the gap between his workflow and the average lawyer's as mainly an adoption problem. A few hours of learning. Download the app, write a detailed prompt, build your first skill. He says it "doesn't require technical skill" in the same piece where he describes building a command-line tool that parses legal documents into spoken audio via an AI voice API, and where his core workflow involves Claude manipulating Word documents at the XML layer. These are not the same universe. He has the technical fluency to build and debug this stuff. Most practicing lawyers do not, and telling them the barrier is low does them a disservice.

The economics get selectively presented too. He notes that three associates would have needed until morning to produce what he did in two hours. Maybe so. But those three associates are also three sets of eyes. The entire architecture of large-firm review — the junior associate's draft, the senior associate's markup, the partner's final pass — exists because self-review has limits that don't disappear when the first draft arrives faster. One lawyer reviewing AI output at 11 PM is efficient. Whether it's sufficient is a different question, and it's the one the piece doesn't engage with.

And the billing discussion opens a trapdoor he walks right over. Subscription pricing works, he says, because AI compresses the marginal cost of delivery. True. But his clients are venture-backed startups — exactly the people most likely to grasp that logic and turn it back on him. If the work that used to take £30,000 of associate time now takes two hours, how long before a savvy founder asks why the subscription isn't a fraction of what it is? The efficiency argument that justifies the model is the same argument that eventually compresses it.

Where Shapiro lands, though — in his closing paragraphs — is where the piece earns its keep. Experienced lawyers, he argues, are sitting on exactly the asset that AI makes more valuable: judgment. The pattern recognition built over a decade of practice. The instinct for which fight matters and which to concede. The contextual read that no model has access to. AI doesn't replace that. It gives it leverage.

That's the real insight, and it deserved to be the thesis rather than the send-off. The professionals who will benefit most from these tools aren't the ones who are best at prompting. They're the ones who have spent years building the judgment that makes the output worth trusting. The prompt is not the skill. You are the skill. The prompt is just how you finally get to scale.