A controlled system for formal legal correspondence.

The operational issue was not simply drafting speed. The harder problem was legal consistency: the same practice needed formal legal opinions, advice notes, client letters, transmittals, execution checklists, fee requests, periodic reports, demand letters, acknowledgements, and internal memoranda to follow the same voice, legal posture, and document discipline.

Generic prompting did not solve that. It produced plausible text, but it did not reliably preserve firm voice, apply standing review positions, assign references, or distinguish an internal discussion from a dispatch-ready document.

The workflow is a set of specialised skills and reference files rather than a single prompt. That distinction matters: a legal opinion is not just generated text - it is the firm's position, based on source material, expressed in a controlled form. The system separates the task into discrete stages, each with its own contract, so that nothing about the document is produced incidentally.

The point was to encode drafting judgment as a workflow, not to encode generic legal prose into a model.

Each document type has a known structure. External documents use a controlled register, fixed terminology, and disciplined closings appropriate to formal correspondence. Internal discussion stays direct, because the reviewer needs advice rather than a letter.

The final gate is still human. The system runs an automated finality check against a defined set of criteria before the document reaches the reviewer, but the reviewer decides whether the advice should be dispatched. The check exists to remove the routine failure modes from the reviewer's attention, not to substitute for it.

Execution-stage documents move into a separate signing workflow, described in the document execution case study. That workflow treats finality, signer authority, attachment completeness, and audit trail as its own control problem.

The next step is stronger evaluation. The useful measure is not whether the text sounds legal. It is whether partner edits reduce over time, whether issue spotting improves across document types, and whether the finality gate catches the same problems a senior reviewer would catch.

This page describes the architecture. The private implementation - the templates, the standing positions, the validation rules - stays with the firm. The portfolio is the demonstration; the work is where it happens.