Meeting minutes from raw transcripts: Instructions & setup notes
How to install and use the Meeting Minutes Prompt System
This document explains how to set up the two agents, how to load the knowledge base, and how to run the system correctly in practice.
The system is intentionally manual.
This is not a limitation - it’s what makes it reliable and controllable.
Installing the agents
You will create two separate agents inside your LLM of choice (for example ChatGPT, Claude, Gemini, or similar tools that support long prompts and file uploads).
Each agent has:
its own instructions (prompt)
its own role
access to the same knowledge base
They must remain separate.
Agent 1 - Enrichment & Rewriting
Purpose
This agent processes one chunk of transcript at a time.
Its job is to rewrite, correct, and enrich the text locally, without performing any global synthesis.
How to create it
Open your LLM and create a new agent (or start a new dedicated conversation).
Paste the full Agent 1 instructions exactly as provided.
Save the agent or keep the conversation dedicated to this role.
Upload the knowledge base file (Word or PDF).
Upload any reference documents related to the meeting topic.
This agent will be reused for every transcript chunk.
Agent 2 - Final Consolidation
Purpose
This agent works on the full document, after all chunks have been processed by Agent 1 and manually merged.
Its job is to:
deduplicate content
harmonize terminology and style
produce the final, coherent meeting minutes
How to create it
Create a second, separate agent or conversation.
Paste the full Agent 2 instructions exactly as provided.
Upload the same knowledge base file used for Agent 1.
Upload the full merged document produced from Agent 1 outputs.
Use this agent only when a global revision is needed.
Preparing the knowledge base
The knowledge base defines structure, style, tone, and rules for the final document.
How to prepare it
Use the provided knowledge base template
Save it as a Word or PDF file
Replace example text with meeting-specific rules where indicated
Do not modify parts of the template that are not marked as examples
Important
The same knowledge base is used by both agents
Any change here affects the entire system output
Think of the knowledge base as the configuration file of the system.
How the system works (recap)
At a high level, the workflow is:
Obtain a meeting transcript or written notes
Split the text manually into chunks of up to ~2,000 words
Run Agent 1 on each chunk
Copy and paste each rewritten section into a single document
(Optional) Run Agent 2 on the full document to finalize it
Each step is explicit by design.
Practical usage notes
Chunk size matters
Do not exceed ~2,000 words per chunk
Larger chunks reduce accuracy and control
Smaller chunks are fine if the content is dense or complex
Manual merging is intentional
Do not ask the LLM to merge chunks automatically.
Manual copy/paste:
preserves ordering
avoids hallucinated joins
keeps full control over the document
When one agent is enough
You do not always need Agent 2.
Using only Agent 1 is often sufficient when:
the meeting is short
accuracy is more important than synthesis
repetitions are minimal
you want fast, clean output without global restructuring
Agent 2 is for refinement, not necessity.
Common mistakes to avoid
Uploading the entire transcript at once
Letting one agent perform both local rewriting and global synthesis
Skipping the knowledge base
Treating the system as “automatic”
This system works best when each step does one job only.
Final suggestions
Keep agents simple and focused
Treat prompts as configuration, not conversation
Update the knowledge base before changing prompts
If results degrade, reduce chunk size first
This system is designed to be:
understandable
debuggable
adaptable over time
It behaves much more like a manual tool than a black box.