Case Study: The Document Writer Agent

Throughout my time leading companies, I have consistently wrestled with frustrations regarding written documentation. The pain points are almost always the same:

• Inconsistent Structure: Recurring documents change format every time they are written.

• Ignored Templates: Templates exist, but they are rarely used or are applied haphazardly.

• Outdated Tools: The templates themselves are rarely kept up to date.

• Bad Data: Content is often not drawn from the correct source (usually copy-pasted from a previous, outdated document).

• Undocumented Work: Simply put, too few documents are created, leaving critical information unrecorded.

• Time Drain: When documents are created, the process takes far too long.

So—how can AI help here?

To solve these challenges, we built the “Document Writer Agent.” This agent is a single Large Language Model (LLM) supported by a robust architecture of features designed to streamline the writing process.

The Foundation: Intelligent Templates

The first challenge was fixing the templates. We switched to simple markdown files containing titles, main headings, and generic instructions on what to include.

These files are compatible with almost any system, and crucially, LLMs process them incredibly well. The Agent doesn't just look at the template; it has access to key company systems, including vectorized documentation (internal know-how) and our CRM, ERP, and Finance systems.

The User Experience: Removing Friction

The second challenge was adoption. To get people to actually use the templates, the process had to be easier than the alternative.

We built an app that adapts to the user’s needs, starting with a simple template selection. From there, the user chooses between two modes: Manual Input or Guided Input.

1. Manual Input

This mode is designed for speed. It offers a text interface where you write a condensed, rough version of the document. The LLM then takes your notes, structures them according to the template, and exports the result to a Google Doc (or any online word editor).

This interface is exclusive to the app and is perfect for users who want to generate the framework quickly and add the bulk of the content themselves later.

2. Guided Input

This is where the app really shines. The AI uses the template to drive a conversation, asking specific questions about each section of the document. You engage in a dialogue with the LLM, writing and maturing the document together.

To make this accessible, we made Guided Input available through three different channels:

• In-App Chat: A standard chat window within the application.

• Slack Integration: Using the command /document-writer starts a thread where you can chat with the agent directly in Slack—one less interface to worry about.

• Voice Conversation: Powered by ElevenLabs, this allows for a real-time spoken conversation with the LLM.

The "Commute" Advantage

My personal favorite is the Voice Conversation feature. It transforms the writing process into a natural thinking process. You can talk, take a step back, adjust your thoughts, and ask for advice in real-time.

Suddenly, boring documentation is being written while I’m driving to the cabin or commuting. I can finish critical work without ever looking at a screen.

Current Capabilities

The agent is currently producing high-quality versions of:

• PRINCE2 Management Products: Including Project Briefs and Stage Reports.

• Decision Notes: For the board or management teams.

• Business Plans.

• Monthly Reports: Covering both financial and operational data.

How was it made?

The agent is part of the Brage AI Hub software where the front end was designed in Google AI studio, stored in Github and then the backend was made in Google's Anti Gravity agentic IDE. Currently running locally to ensure local LLM support through Ollama is present (all agents have a Privacy button forcing local LLM usage). Can be spun up on Cloude Run via integrated tools for more wide spread deployment, or, on an on-prem server having minimum 32GB unified or GPU memory allowing for up to 32B parameter sized models (leaving some room for other things running) - the model is running with Gemini 3.0 Flash/Pro depending on use with default local model being Open AI GPT-OSS 20B with a very good Norwegian alternative in Normistral 7B IT.

The Document Writer Agent is already saving me hours of work while simultaneously improving the quality of our documentation. Better yet, other AI agents can utilize it as well—but I’ll cover that in a future article.