Advice Documents Drafted From the Adviser's Own Templates

The challenge

Every advice business has its own house style. One practice wants a file note in bullet points with an action list at the end. Another wants flowing paragraphs with direct quotes from the client. A single "summarise this meeting" prompt produces something generic that still needs rewriting, which defeats the purpose. Advisers also needed confidence that a draft would never invent figures or commitments that were not discussed.

What we built

A template builder and a two-pass drafting engine for FinPlan:

• A four-step template builder: choose the document type, arrange its sections, set the tone and level of detail, then review. Sections drag into order, and each has its own instructions, length and output format: bullets, numbered list, paragraphs, table or checklist.
• A starting library: 20 document types across wealth, insurance and mortgage advice, including file notes, fact finds, statements and records of advice, investment policy statements, retirement projections, aged care, SMSF investment strategies and insurance needs analyses. 25 ready-made section sets and 30 instruction presets, such as compliance focused or action oriented, give advisers a head start.
• Settings that become instructions: target audience, tone, level of detail and whether to quote the client are written into the prompt alongside fixed guidelines. Never fabricate. Keep figures exact. Account for every speaker. Leave out small talk. Section-level instructions take priority over the global settings.
• Two passes for custom instructions: the first pass drafts the report from the template. If the adviser has added instructions of their own, a second pass applies them to the finished draft and changes only what those instructions require. If the second pass fails or times out, the adviser gets the first draft.
• Long meetings handled: when a transcript is too long for the model, it is trimmed at a paragraph or sentence boundary and marked as shortened, so the model knows it is working from part of the conversation.
• Clean output: drafts come back as Markdown and pass through formatting rules and a validator before they reach the editor.

Technology

GPT-4o-mini drafts the reports, chosen for speed and cost, with the model, temperature and timeouts held in configuration. Kotlin and Spring Boot services build each prompt from the stored template, and templates and reports live in MongoDB. The template builder is a Next.js and React wizard with drag-and-drop sections. Transcripts arrive already cleaned by Readable Meeting Transcripts That Keep the Detail. We run drafting in production as part of An AI Platform That Turns Client Meetings Into Finished Reports.