Method & trust
A confidently-wrong legal answer is worse than no tool at all. Everything about how these skills are built follows from that one sentence.
The trust model
The judgment layer is AI-drafted and expert-reviewed — the model is good at how to approach a task, and review catches tone and gaps. The authoritative layer is sourced from real law and expert-verified — because that is where a hallucinated fact would live, and the judgment layer’s instruction to “sound authoritative and cite sources” would make that wrong fact more damaging, not less.
How to approach the task, the way an election expert would.
AI-drafted · expert-reviewed
The actual law — the facts the judgment grounds itself in.
Sourced from real law · expert-verified
Make the model approach the task the way an expert would, even when the facts still come from the model. Achievable today; polished, but possibly wrong on hard legal specifics.
Good enough alone for method-dominant tasks, where exact local law isn’t make-or-break.
Connect the model to verified, current, jurisdiction-specific facts. This is the real moat — and it is not achievable by instructions alone. It needs carried references or runtime retrieval.
Required for fact-dominant tasks like cure deadlines or records releasability — or they become liabilities.
A loop drafts, checks against explicit criteria, revises, and converges. It’s worth it only when the verification step is more reliable than the generation step — the model can be a mediocre generator but a good checker.
Strong fit
Weak fit
A loop is only as good as its test point. We build loops only where a real verification signal exists — self-consistency is not correctness.
Production model
Scope the task & jurisdiction
Pick a language-heavy, judgment-driven task that consumes officials’ time, and decide what jurisdiction the facts must be true to.
AI drafts the judgment layer
The model writes the thin, agnostic SKILL.md — how to approach the task like an expert. AI is genuinely good at this part.
Source the authoritative layer
Every fact is pulled from real statute or guidance and cited. Never AI-generated from memory. This is the moat and the risk, so it gets the strictest standard.
Expert review
A subject-matter expert reviews tone, completeness, and — above all — the legal facts. The SME is the trust gate. The pipeline cannot skip them.
Adversarial testing
The expert tests the skill in a real tool against real questions — careful ones and terse, messy, real-world ones — to see whether it holds up under pressure.
Version, date & publish
The skill carries the date its law was verified, so staleness is always visible. Then it ships.
Non-negotiable gates
SMEs are the trust gate. The plan cannot skip them.
Deliberately out of scope