@cartainc created a system prompt for its agent by using Lucidchart to diagram a domain expert’s full workflow. Then an internal prompt generator converted that into JSON, which was fed into the agent programmatically. “Teams that best understand the domain are contributing directly to prompt and context design, improving velocity and relevance without needing to train models or learn ML,” says @JayantTikmani. Here's their gold standard prompting techniques, which rely heavily on in-context learning. For the full story of how Carta built an agent that saved internal teams 3,500+ hours per month, follow the link in the comments.