Client Meeting: Software Development Issues

 Client meeting 5/15/26, notes taken by Calendly

Summary The meeting centered on the challenges and philosophies of structured software development, particularly in the context of integrating AI and agentic systems. Client and Michael McCafferty discussed the limitations of AI in software architecture versus its strengths in code generation, emphasizing the persistent 'knowing-doing gap' in both human and AI-driven development. They explored best practices, such as the Lean Startup model, advocating for early and continuous user involvement to ensure software meets real needs and garners user buy-in. The conversation delved into the technical and philosophical aspects of teaching AI how to think, including the use of retrieval augmented generation (RAG) to manage context limitations in large language models. They also addressed the importance of structuring both the development process and personal life, with actionable suggestions like maintaining checklists and aligning life plans with work objectives. The meeting concluded with plans to review checklist data in the next session and a mutual commitment to ongoing collaboration and self-improvement.

Action items Client   • Next week, review and analyze the checklist data collected so far, and meet with Michael McCafferty to discuss insights and gaps between the life plan and current measurements.

Discussion AI and Software Development Capabilities and Limitations The participants discussed the strengths and weaknesses of AI in software development, emphasizing that while AI is highly effective at writing code within well-defined frameworks, it struggles with high-level architecture and design. Both agreed that AI lacks the ability to perform architectural work, and that structured software development principles remain essential regardless of AI involvement. The conversation highlighted the 'knowing-doing gap'—the difference between knowing best practices and actually implementing them—attributing this gap to human tendencies to seek immediate gratification from coding rather than following structured processes. Bridging this gap was identified as a key challenge for both AI and human developers. Software Development Methodologies Waterfall vs. Lean Startup The meeting compared traditional waterfall and lean startup methodologies, noting that waterfall involves building software before seeking user adoption, while lean startup emphasizes early user engagement and building to their specifications. Michael McCafferty advocated for the lean startup model, stressing the importance of understanding both the needs of project funders and end users, and involving users early to ensure buy-in and reduce resistance. The discussion also highlighted the value of rapid prototyping and ongoing user involvement to ensure the final product meets real needs. Challenges with Legacy Systems and COBOL The conversation explored the difficulties of working with legacy systems, particularly those written in COBOL. Client shared an example of integrating Notion with a hotel scheduling system in COBOL, highlighting the scarcity of training data for AI models due to proprietary codebases. Michael McCafferty, with extensive COBOL experience, noted that while COBOL is structured and English-like, its diversity makes it hard to train AI models, and traditional human programmers are still needed for COBOL work. They agreed there is a business opportunity in developing AI models that can handle COBOL, but neither participant is eager to pursue it. Best Practices and Structuring Agentic Development Client discussed curating a library of best practices for agentic (AI-driven) development, emphasizing the opinionated nature of software development and the need to break down best practices into atomic steps with checklists for each phase. Michael McCafferty recommended adopting the lean startup model and starting with understanding the needs of both funders and users, advocating for top-down design that begins with business goals and user needs rather than technical details. The discussion included the importance of constraining AI development with initial design and philosophy, and debated how much AI should mirror human thinking versus following structured approaches. Technical Constraints of Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG) The participants examined the technical limitations of LLMs, particularly their limited context windows and the challenge of teaching them to think stepwise. Client explained that LLMs can only process a limited amount of information at once, making it necessary to clear context and focus on specific tasks. Michael McCafferty suggested using Retrieval-Augmented Generation (RAG) to avoid overloading the context window. They discussed breaking problems into smaller components and guiding the model through sequential steps, as LLMs cannot autonomously manage complex, multi-step reasoning. The conversation highlighted the need for careful orchestration and structuring of tasks for LLMs to be effective. Philosophy of Work, Life Balance, and Personal Development A significant portion of the meeting focused on personal development, work-life balance, and the philosophy behind both. Michael McCafferty reflected on his own career, noting that he was consumed by work in his 20s and neglected personal life, and advised Client to avoid this trap by thinking about life goals beyond work. The discussion included the value of checklists for both professional and personal growth, the importance of aligning daily actions with broader life plans, and the need to periodically reassess one's direction. They also discussed the benefits of seeking experiences outside work, using tools like calendars to schedule activities, and the role of reading and learning in personal growth, while acknowledging the challenges of motivation and the tendency to default to work for gratification. Meta-Philosophy Teaching AI How to Think and the Role of Philosophy in Agentic Systems The meeting delved into the meta-philosophy of agentic AI systems, with Client raising the challenge of teaching AI not just what to do, but how to think, given the limitations of current models. Michael McCafferty suggested that the way AI 'thinks' is shaped by how developers prompt and structure it, cautioning that AI will reflect the developer's own thinking patterns. They agreed that the real challenge is philosophical, not technical, and recommended seeking insights from leading AI companies and their developers. The discussion emphasized the importance of defining the philosophical framework for agentic systems and the need to look beyond technical solutions to broader questions of behavior and structure.