For Lawyers · Published Free
How this firm actually runs — the systems, the workflows, the software, and the mistakes. Written for solos and small firms whose bottleneck is hours, not talent.
Start Here
4 articles·56 min· read in order
Essay
For ten years, running a firm meant duct-taping together SaaS tools that mostly talked to each other. That's ending. Five ideas I picked up before I was a lawyer decide whether you can build your own.
14 min
Essay
Strip off the interface and a practice management system is a relational database, a set of views, and an API. That used to be worth every dollar. Now the database is a weekend, the connectors reach anything, and the real question is which parts you should still rent.
16 min
Essay
Everything in this section — the wiki, the agent briefs, the checklists, the eval set — lives in one directory of plain text. That's what lets you change models on a Tuesday afternoon, including to one running on the machine under your desk.
12 min
Essay
AI can draft, research, and summarize. It can't tell you what you were doing. A working lawyer on what it costs to keep six AI sessions straight in your head — and the kanban board I built to carry the load instead.
14 min
Latest
Essay
On September 29 the President told the executive branch to say 'Super Intelligence' instead of 'artificial intelligence' — then defined the new term as the old statutory definition. It binds federal agencies and nobody else. Your court rules, your Minnesota statutes, and your contracts don't move. Your search terms, your cross-references, and your calendar for November 28 do.
Read the articleOne email when something new is published here. No course, no upsell — the Lab stays free either way.
The Archive
Workflow
Put the last two essays together and you get one design: a model on the desk strips out identity, the token map never leaves the office, only a scrubbed slice goes to a frontier model for the hard reasoning, and the answer gets re-identified locally before a person reads it. Here's the pipeline step by step, working code for the mechanical half, the seven ways it breaks, the check for each, and what it costs in Comment [17] terms.
Essay
No Minnesota rule says you have to anonymize client data before it goes into a model. The word isn't even in Rule 1.6. What Comment [4] gives you is the test a scrubbed prompt has to pass — no reasonable likelihood anyone can figure out the client or the situation involved — and the half lawyers miss is the situation, not the name. Where Minnesota is looser than the ABA, and where nothing takes the place of the client's signature.
Essay
On August 25 Apple wrote 'without counting tokens or worrying about rising cloud costs' into its own press release and priced the machines at $899 to $5,499. They started arriving September 22. What that changes for a small firm, what it doesn't, the footnote Apple hopes you skip, and what I'd buy this quarter — after you test, not before.
Essay
A sentence in Minn. Stat. § 325M.20(a) expired on schedule on January 31, 2026. Since that day the Attorney General can sue a Minnesota business with no letter and no cure period, at $7,500 per violation plus the state's litigation expenses. Every 2024 and 2025 client alert describing a Minnesota cure period is now wrong — and the Legislature wrote the reason to bring this work to a lawyer, not a consultant, right into § 325M.18(f).
Essay
Everybody's pitch is the same: Minnesota should go win an AI consortium the way Austin won MCC and SEMATECH. Both of those died or moved out. What left Austin was the part Texas bought. What stayed was the part built into people and institutions. Assurance, being able to prove an AI system does what its seller says, is all the second kind.
Essay
Here's a claim you can test: a downloaded model checkpoint is an untrusted executable, not a document. And the three-year Hugging Face record isn't one hack. It's an escalation — a code-execution exploit, a scanner bolted on top of it, a way around that scanner, and finally a backdoor one layer up that no scanner reaches. Patching a layer doesn't close the category.
Essay
The paperclip maximizer is the famous frame and the wrong one to lead with. It has taken serious, published, recent criticism, and it pushes the problem off onto a superintelligence that doesn't exist. The real unsolved problem is here now: we can prove narrow things about small networks and nothing broad about large ones, and the gap between what a vendor intended and what anyone can prove is exactly where a negligence claim gets built.
Essay
Two claims you can check me on later. The frontier model itself, not just the trailing tier, is a depreciating asset whose durable value rounds to zero. And the campaign to restrict open-weight AI is best understood as incumbents reaching for the only moat that doesn't melt: a regulatory one. Lawyers, of all people, should recognize the costume.
Workflow
The most important decision in an AI-assisted practice isn't which model you use. It's that agents produce drafts and staged actions while a human fires everything that crosses the firm's boundary. That rule isn't training wheels you grow out of. It's a permanent security model, and it happens to make the supervision duties workable.
Workflow
Over two days this firm published thirty verified articles on Minnesota law, and eight of the legal premises I handed down in the assignment briefs turned out to be false. That's a lesson about delegating — to agents and to associates — because the senior lawyer's memory is the least reliable part of the pipeline.
Essay
A solo firm can now verify law like a research department, and the reason is a handful of nonprofits and public offices that almost nobody in this profession pays. Here's what that infrastructure actually is, the four places it's already breaking, and why a donation line item in your budget isn't charity. It's a maintenance bill.
Essay
A legal research vendor offered to put an AI assistant inside their platform. That's backwards. The assistant is where the work happens now, and the research service is the thing that has to come to it. What that means for the renewal notice every firm is about to get.
Essay
The reader of a law firm's website isn't just a person typing into Google anymore. It's an assistant pulling a passage to answer somebody's legal question, and that reader rewards depth, verbatim quotation, pinned authority, and honest statements of what a statute doesn't say. That happens to be the same list as writing well.
Tool
A small firm doesn't need a database server. Picking one means signing up for operational work that produces nothing for a client. One file on your own machine, with WAL, foreign keys, FTS5, and a real migration path, can do more than most lawyers realize.
Workflow
Getting a script to write a .docx or fill a PDF is easy. Getting one that opens right on the other person's machine, prints right, and doesn't embarrass you in front of a clerk means beating five failure modes — and the code that made the file can't see any of them.
Essay
Everything in this section — the wiki, the agent briefs, the checklists, the eval set — lives in one directory of plain text. That's what lets you change models on a Tuesday afternoon, including to one running on the machine under your desk.
Workflow
My pre-publication link check, a shell loop, told me 'Safe to publish' for weeks without checking a thing. The bug, why zsh hid it, the fix, and the thirty-second habit that catches the whole family of bugs like it.
Tool
The stack I'd actually install to run Qwen on firm hardware — models, runtimes, quantization, what fits in how much memory — and the tier election it lets you hand the client. If the client picks the tier, the routing has to enforce it.
Essay
My prediction, with the failure conditions attached: open weights push the value of the model itself toward zero, and the money moves to whoever owns compute — and under that, power. And what a law firm should do differently if I'm right.
Essay
Strip off the interface and a practice management system is a relational database, a set of views, and an API. That used to be worth every dollar. Now the database is a weekend, the connectors reach anything, and the real question is which parts you should still rent.
Essay
For ten years, running a firm meant duct-taping together SaaS tools that mostly talked to each other. That's ending. Five ideas I picked up before I was a lawyer decide whether you can build your own.
Workflow
The most common citation defect in AI-assisted legal writing isn't the fake case. It's the real quotation that stops one clause too soon, and you'll almost never spot it on review. Here's a detector that catches it mechanically.
Essay
Most legal futurism can't be proven wrong, which is why it's always half-right and never useful. Twelve predictions about how AI reshapes legal practice — each with what drives it and the specific evidence that would prove me wrong.
Tool
If your AI tool reads statutes through a summarizing fetch layer, you don't have a legal research system. You have a paraphrase generator with citations stapled on. Here's the scraper I replaced it with, and the four bugs it caught.
Tool
Nobody in law builds test sets, so nobody can tell a workflow that got better from one that quietly got worse. How to build ground truth from closed files, which numbers matter for which job, and why cost-weighted error is the only honest score in a law office.
Essay
Legal aid can't hire against one-year grants, and volunteer hours eat up the supervision they're supposed to add. A forward purchase commitment, tagged by specialty, fixes both — and bundles in the one thing big firms actually have to give.
Tool
Why dumping a matter file into a context window fails, why embedding search stumbles on legal text, and how to build retrieval that finds the right paragraph and tells you exactly where it came from.
Essay
Every system that ties work to a matter on its own gets tempted to always come up with an answer. In a law office that's the wrong design, and Rule 1.5 is why.
Workflow
Twelve calls, four web inquiries, a portal message and three e-service notices isn't a lot of information. It's a lot of interruptions. How to sort inbound automatically, route it by what it demands from you, and get comfortable not seeing the stuff that doesn't matter.
Tool
The strongest confidentiality posture is client data that never leaves the building. Here is what local models actually cost, what they are genuinely good at, where they fall down, and the hybrid architecture that is usually the right answer.
Workflow
Unbilled time isn't a discipline problem. It's a capture problem: the work happened somewhere your billing system couldn't see. How to rebuild it honestly, why rounding decides whether approved equals billed, and the one rule that never bends.
Essay
The dangerous bug in a firm's own software is never the crash. It's the button that does nothing, the file saved to a default folder, the integration that quit syncing in March and told nobody.
Workflow
The mail shows up as a stack of scanned PDFs and leaves as filed documents, calendared deadlines, and a short list of things somebody has to decide. Six steps, and a hard rule about which ones a human has to sign off on.
Essay
Every AI-assisted workflow leaves a trail: prompts, retrieved documents, drafts you threw away, transcripts. Is it work product? Is it privileged? Does a litigation hold reach it? A few trial courts have started answering the first two, and they've come out different ways. Most firms are keeping all of it without ever deciding to.
Tool
There's an open standard for exactly this problem. My practice management vendor doesn't implement it. So the connector is four hundred lines of my own Python, and the interesting part isn't the code. It's why every firm is writing it separately.
Workflow
The software you build for your own firm gets used by lawyers and legal assistants, not engineers. The fastest way to make it useless is to borrow the vocabulary of the tools you built it with.
Workflow
AI gives you generic legal work because it has no idea where it is. A structured Obsidian knowledge base — plain markdown, a real schema, a source on every fact — turns a general assistant into one that already knows your practice.
Essay
Do you have to tell a client you used AI on their matter? What happens to the fee when a two-hour task takes twenty minutes? The disclosure question really is unsettled. The fee question isn't, and it cuts against the lawyer.
Workflow
Minnesota's three-day mail rule moved in 2020. Rule 6.05 was abrogated; the text now lives in Rule 6.01(e). If you learned it the old way, you've been carrying a wrong rule for years — and so has every AI you ask.
Workflow
A fabricated case is the easy failure. It doesn't exist, so somebody notices. The dangerous one is a real case, cited perfectly, standing for something it never said. Here's the four-pass gate that catches it.
Workflow
One general-purpose AI assistant gets you a capable generalist with no limits and no handoffs. No real office runs that way, and yours shouldn't either. How to write agent briefs, and why the limits matter more than the skills.
Workflow
Every firm's real institutional knowledge lives in its form file, and every firm's form file is quietly wrong in places nobody's checked since 2014. Keeping templates clean, how stale rules spread, and why AI should fill in a template, not rebuild one.
Essay
Nine of every ten ethics complaints in Minnesota close with no discipline at all, and a lawyer had to answer every one of them. Here's what all that regulation costs the people who never get a lawyer in the first place.
Essay
AI can draft, research, and summarize. It can't tell you what you were doing. A working lawyer on what it costs to keep six AI sessions straight in your head — and the kanban board I built to carry the load instead.
Nothing matches that.
One email when something new is published here. No course, no upsell, no newsletter about the newsletter — the Practice Lab stays free either way.
Prefer a reader? Subscribe by RSS →
The Premise
Most people with a real legal problem never see a lawyer. Not because they have no claim, and not because no lawyer would take it — but because an hour of legal help costs more than they have. Legal aid is chronically underfunded; pro bono is essential and will never scale to the size of the gap.
If we are serious about opening the courthouse door to the people currently standing outside it, there is one lever with enough leverage to matter: the cost of producing legal work has to come down. A lawyer who can carry more files well is a lawyer who can say yes to files the economics would otherwise rule out.
So everything here is written for other lawyers — solos, small firms, anyone whose bottleneck is hours rather than talent — and everything here is free.
There is no signup, no course, no subscription, and nothing for sale. We are not asking for donations.
If something here saves you time, spend some of it on people who could not otherwise afford you.
Take the case at a rate that works for someone on a fixed income. Keep a sliding scale. Say yes to the file you would otherwise have to turn down on economics. That is the entire price.
Nothing in this section discusses a client, a matter, or a real case. Screenshots use invented clients and file numbers. Everything here is general commentary on practice management — it is not legal advice, it is not ethics advice, and reading it does not create an attorney–client relationship. Every lawyer is responsible for their own compliance in their own jurisdiction.
Corpus Map
Each article is reduced to a TF-IDF vector over its own vocabulary; every pair is scored by cosine similarity, and each article keeps its two strongest links. The layout is a Fruchterman–Reingold force simulation run to convergence at build time — deterministic, so the same corpus always yields the same map. Clusters are emergent, not assigned.
Similarity Matrix
The graph is a projection; this is the object it projects. Cell (i, j) is the cosine similarity between articles i and j. Rows and columns are ordered by agglomerative hierarchical clustering with average linkage over 1 − cosine distance, so the dendrogram's leaf order puts related articles adjacent — the dark blocks along the diagonal are the clusters, not an artefact of ordering.