đź“ť

Writing meeting notes

Last edited time
Apr 9, 2024 11:33 AM
Reading list
Status
1 case
Tools
tags
documentationoperations

Case 1: Write Meeting Notes Like PMP

Configuration

Model: gpt4-turbo-preview

Temperature: 0

Prompt Framework: CO-STAR

Introduction

Project management values detailed and actionable meeting notes. This use case highlights the use of ChatGPT in making meeting notes that follow Project Management Institute (PMI) guidelines. Although tools like Tl;DV provide conversation transcripts, they often lack detail required for project management documentation. Our goal is to fill this gap, ensuring complete meeting notes that align with project management best practices.

Hypothesis

We believe that using ChatGPT to process raw meeting transcripts can significantly enhance project management note creation. This method assumes that a more detailed use of AI, especially through tailored template filling and iterative refinement, can produce thorough meeting summaries that align with PMI best practices. This is an improvement over existing automated solutions like Tl;DV, which often generate abstract and incomplete summaries.

Solution Implementation

We used a step-by-step process to automate project management meeting notes with ChatGPT. This involved preparing a template, filling in information, and refining it.

Step 1: Template Preparation

The first step was to prepare a template following PMI's best practices for project management. The template covers essential meeting elements like objectives, stakeholders, key discussion points, constraints, and lessons learned. We created the template in stages, first making a draft with ChatGPT, then reviewing and adjusting it to suit project management needs.

‣
Meeting Notes Template

Step 2: Incremental Information Filling

We had to take a step-by-step approach to handle the massive task of processing meeting transcripts and extracting relevant details. Trying to handle all of it at once resulted in missed information and incomplete coverage.

Breaking the task down into manageable parts allowed ChatGPT to generate detailed content for each section of the meeting notes. This method captured more detailed information, including crucial nuances often missed when trying to process everything at once. However, this approach also led to some problems, such as repeated or misplaced information, especially in sections like 'Requirements' and 'Constraints'.

Step 3: Iterative Refinement

To address the problems found in the initial filling stage, we used an iterative refinement approach. We made specific prompts for each section of the template to better guide ChatGPT. We used few-shot prompting to make clear the difference between similar sections, such as 'Constraints' and 'Requirements', by giving both good and bad examples. This careful prompting reduced repeated and irrelevant information, making the meeting notes more accurate and focused.

We improved the process by using insights from the research report, particularly to deal with repeated data and to ensure accuracy. Techniques like segmented processing, temperature adjustment, and advanced contextual awareness from the study, were key in improving the automation process.

In short, we used a structured approach to implement this use case, using the abilities of ChatGPT, and improving it through repeated testing. It shows the progress in creating meeting notes, moving from manual collection to advanced, AI-supported automation, while following project management best practices.

Evaluation of Results

We conducted a comparison between similar requests to ChatGPT4 and Tl;DV. It's important to note that Tl;DV doesn't allow for long custom prompts. The results are presented in the table below. From our analysis, we can conclude:

  • ChatGPT responses are better structured and can be formatted according to user preferences.
  • ChatGPT generates more comprehensive responses.
  • Unlike Tl;DV, which sometimes fills certain sections with irrelevant information, ChatGPT can indicate when no relevant information is found.
‣
Comparison of results

Limitations and Constraints

  • ChatGPT sometimes has issues with repeated information and incorrect placement of information across different sections of the notes. For example, it can have difficulty distinguishing between project management terms like requirements and constraints.
  • ChatGPT struggles to maintain context and handle large data inputs, leading to potential errors or inaccuracies when dealing with large volumes of information. There are instances where the context length reaches 6200 tokens, which could result in a loss of up to 20% of the facts.
image
  • There is a need for human oversight to ensure the AI-generated notes are accurate, especially in correctly distinguishing project management categories and accurately reflecting the meeting's content and decisions.

Conclusion and Next Steps

Using ChatGPT to automate project management meeting notes has given us valuable insights into improving meeting documentation. By following PMI recommendations and using a structured approach to fill out meeting note templates, we saw an improvement over alternative AI summarization tools like Tl;DV. Our method, based on repeated refinement and specific template use, matches well with top project management practices, offering a clearer and more useful process.

  • Process Automation: This involves the automated retrieval of meeting transcriptions and the execution of iterative prompts. It also includes adding chains of agents to validate and refine outputs.
  • Including user feedback in iteration cycles: Regularly incorporating user feedback can ensure the system continues to meet user needs and expectations.
  • Refining AI model prompts: Further enhancement of AI prompts can improve the quality and accuracy of generated meeting notes.
  • Improving Context Management: Enhancing context management through Retrieval Augmented Generation (RAG) can lead to improved performance and greater consistency with existing project management documentation.
  • Implementing Comprehensive Automatic Identification: Develop features that can not only identify and assign tasks but also categorize other items such as requirements, risks, and lessons learned into appropriate registers, enhancing overall project management efficiency.

Final Prompt

‣
Expand

Example of notes