The one idea
Ask a frontier model to run your practice end-to-end and it will produce something fluent, occasionally wrong, and impossible to audit. Do the opposite — build a narrow, inspectable operating system around the model — and you get something a lawyer can actually rely on and file.
What is the method, in one sentence?
Let the model propose meaning; let deterministic software decide what is allowed; let registered workflows perform the work under validation; and let a qualified human approve anything consequential. The firm’s durable advantage is that layer of skills, workflows, retrieval, evaluation, and guardrails — not the underlying model, which any competitor can rent too.
That layer is portable across model vendors, survives any single model’s retirement, and is fully owned by the firm. It is the answer to the question every managing partner should ask: if the model is a commodity everyone can buy, where is my advantage? Your advantage is your decade of filed work product and the way your firm uses a model on it — kept inside your building.
Four layers
The system has four seams, each answering a different question.
- Skills are compact operating manuals that tell the model when a class of work applies and what rules to load — a discovery summary is not a settlement draft and should not carry the same instructions.
- Routing turns a plain-language request into a typed, registered intent, so the machine dispatches a known workflow rather than improvising one.
- Workflows execute bounded, ordered steps under machine-enforced contracts, producing artifacts, validation, and an audit trail.
- Goal orchestration decomposes an outcome too large for one workflow into independent workstreams, then verifies and stops at a defined finish line.
The strength of the system comes less from any single prompt than from the seams between interpretation, policy, execution, evidence, and human judgment.
Why this is different from “prompt engineering”
A clever prompt is a party trick; it works until the input changes. A skill with a discriminating description, a typed workflow with a stop-on-failure runner, and a retrieval layer grounded in your own filings is software — versioned, tested, and promoted only on evidence. The difference between the two is the difference between a demo and a system a firm can put its name on.
The parts that keep a lawyer safe
Three properties of the architecture matter more than raw capability:
Generation is separated from assembly. The model writes the words; deterministic code owns templates, formatting, citations, and file construction. That makes output easy to review and assembly easy to test.
Facts are separated from style. The firm’s past work teaches voice and structure; it never supplies the facts of a new matter. Missing facts are flagged as gaps, not invented.
Quality-critical steps fail loudly. When a check cannot run — a source is missing, a validator is down — the build stops and says so, rather than shipping a confident guess into a court file.
Why it is teachable
Because the method is architecture, not secret sauce, it generalizes. The same pattern that decomposes a family-court financial trace decomposes a contract abstraction or a regulatory memo. PLI breaks the method into modules — skill authoring, workflow design, agent orchestration, retrieval, evaluation, guardrails — with generic exercises and decision checklists, and teaches a firm’s own people to run and extend it. The goal of an engagement is a firm that is more independent, not more dependent.
The rest of this hub takes the layers one at a time. Start with skills, not chatbots, then deterministic document generation, retrieval over your own work product, and evaluating legal AI without fooling yourself.
This article is general information from a technology consultancy, not legal advice, and does not create an attorney-client relationship. Figures describing the founder’s own practice are illustrative, not a promise of results.