Method
Find the ambiguity before the court does.
A drafter's system for Massachusetts legislation, built as a set of Claude skills over two years and calibrated against what the Supreme Judicial Court has actually done with ambiguous text.
The premise
A court reads a statute the way nobody in the room read it: cold, years later, with a dispute already in hand.
In House Counsel you see the whole arc — a bill drafted, amended on the floor, engrossed, enacted, and then construed by a court working only from the text. That vantage teaches two things. First, most interpretive fights are visible in the draft to anyone who knows what to look for. Second, some of the ambiguity is there on purpose. Legislation passes on language that several factions can read their own way, and no drafter's tool should pretend otherwise.
So the system is built as decision support, not quality control. It does not tell a sponsor to remove an ambiguity. It tells the sponsor what interpretive options the ambiguity creates, who bears the cost of the uncertainty, what litigation risk attaches to each reading, and whether the ambiguity serves or undermines the bill's own purpose. The sponsor decides. The system records the decision and the reason.
The anchor for that framing is the Supreme Judicial Court's own: in Camargo's Case, 479 Mass. 492 (2018), the court treated cross-chapter definitional differences as intentional legislative design — “it is up to the Legislature to decide how much uniformity to impose.” A tool that flagged every such difference as an error would be wrong about the law.
Four stages, in order
The order matters. Each stage assumes the one before it has run, and the cheapest checks come first because a live structural seam corrupts everything downstream.
-
Lint the seams
A deterministic scan for three patterns with a long litigation record: relational phrases (“in connection with,” “relating to,” “with respect to”), bare time references (“the first time,” “initially”), and unclear cross-statutory references (“notwithstanding any other provision of law,” “as otherwise provided by law”). It is not a classifier and it generates nothing. It flags the seam and gates it.
-
Detect the ambiguity — Application 1
Three detection layers. STAT finds terms the General Laws define differently across chapters — “employee” under c. 149, c. 151A and c. 152 are three different tests, and a bill that uses the word without choosing has chosen none of them. REG finds regulatory overlays that have narrowed or shifted a statutory term. UNDEF finds terms the bill leaves undefined that are defined somewhere else in the code — the definition a court will borrow whether the drafter meant it or not.
Every alert carries a severity and gets a disposition: add a definition, cross-reference an existing one, accept the ambiguity with the risk stated, or dismiss with a coded reason. A critical alert the sponsor wants to keep goes through a Challenge Protocol — a devil's-advocate analysis, a failure-mode inventory, a distributional stress test, and a written response of at least 150 words. The friction is the point.
-
Simulate the fight — Application 2
For an ambiguity the sponsor accepts, or any provision a client asks about: generate the competing readings each side would argue, name the canons of construction that support each, predict how the Supreme Judicial Court would resolve it and with what confidence, and draft the language that would have made the question go away. The predictions are calibrated against a corpus of Massachusetts appellate statutory-interpretation decisions, coded for which canons the court actually used, in what order, and how firmly it resolved the question.
-
Score the fix — Application 3
When the resolution is a new definition, the definition itself gets tested: scored from 0 to 100 across coverage, precision, clarity and conflict; checked against twenty-seven known drafting defects; compared against roughly 1,800 definitions already in the General Laws and the Code of Massachusetts Regulations. An intentional delegation to an agency is honored as such, not penalized as vagueness.
What the system does not do
Pattern detection finds the words that are there. It does not find the sentence that is missing.
The most expensive defects in the casebook are silences — a statute that says nothing about waiver, nothing about retroactivity, nothing about which of two definitions governs. No scanner catches an omission. The adversarial simulation does, sometimes, because asking “what will each side argue?” surfaces the gap each side will argue about. But the honest answer is that the omission is where the drafter's judgment is irreplaceable, and the system's job is to clear away everything else so that judgment has room to work.
Nothing here is legal advice, and no output leaves the system without a lawyer reading it. The tools are educational; the analysis a client receives is mine.
How it is built
The system is a suite of forty-four Claude skills — structured instruction files that tell the model what to look for, in what order, against which authorities, and what form the output must take. They run in Claude Code and in Claude's desktop application, and they call out to the sources a drafter needs: the Massachusetts Legislature's API for bill and session-law text, CourtListener for appellate opinions, and the drafting and style manuals the General Court and the Supreme Judicial Court publish.
Two design rules run through all of it. Verify before you style. Every claim about a quoted text is checked against the text before any prose is written around it; a confident verdict built on a misread spends credibility that does not come back. Record the decision. Every disposition, dismissal reason, bypass, and challenge is written to an analysis record, so a successor who was never in the room can see what was decided and why.
The public tools are the parts of this that stand on their own: the ambiguity detector as a web demo, the drafting and style manuals made answerable, the tax code drawn as a diagram, a corpus of twenty-five special acts with a query tool in front of it.
One skill runs as an engagement rather than a tool. The Bias Spectrum Audit takes a statute or bill and writes it four ways — the old openly exclusionary version, the rigged version a skilled drafter could write today, today's law, and the most protective version — so that the structure of today's text becomes visible by contrast. Definitional ambiguity is often a "who decides?" decision in disguise, which is why the audit pairs with Application 1 rather than replacing it. Five audits are published; the method is described on that page.