Srihari Sriraman shares insights from his experience building a context-viewer using large language models (LLMs). Initially, he struggled with lengthy 300-word prompts that yielded poor results. By refining his approach to use 15-word prompts, he found a more effective way to communicate with the LLMs. Key lessons included understanding the model's strengths and limitations, and adapting problems to fit the model's capabilities. Sriraman faced two main challenges: segmentation and categorization. Segmentation involved breaking down complex user inputs into manageable parts. He illustrated this with an example message that included various specifications for a product requirements document (PRD). The message contained detailed sections about the problem statement, customer identification, goals, and functional requirements. For categorization, he needed to define the data model clearly, including person records and changesets, ensuring that the system could efficiently handle user data and historical changes. He outlined the requirements for the MVP (Minimum Viable Product), emphasizing functional aspects like user authentication, data entry, and analytics. The design principles prioritized quick data entry and clear visibility into changes. Sriraman also discussed risks, such as ensuring import accuracy, avoiding duplicate records, and managing OAuth configurations. Acceptance tests were established to validate the system, ensuring it met the necessary operational standards. Throughout the article, Sriraman highlights the importance of engineering around the capabilities of LLMs while maintaining a structured approach to problem-solving.