Brad Menezes, CEO of civic tech start-up Superblocks, might have just found one of the best opportunities to innovate with artificial intelligence. He believes that the next wave of billion-dollar startup ideas lies hidden within system prompts used by established unicorn AI startups. These system prompts, which set the behavior of foundational models from companies like OpenAI and Anthropic, are critical in producing application-level AI products. Even more than his photos though, Menezes’s candid observations have really resonated with the public. His tweetstorm on the topic went viral, hitting almost two million views and catching the eye of heavy hitters in Silicon Valley.
Menezes emphasizes three key components of system prompts that warrant examination: role prompting, contextual prompting, and tool use. These prompts instruct AI models to complete very specific tasks and are extremely complex, sometimes even well over 5,000-6,000 words long. Though complex, these prompts are designed in plain language, which both makes them approachable and yet highly refined.
As Menezes points out, the lessons learned by analyzing other people’s system prompts are an excellent introduction to the art of prompt engineering. For instance, the system prompt from Devin begins with a clear directive: “You are Devin, a software engineer using a real computer operating system.” This introduction is exemplary for what the model is able to do and what it’s best suited for – deep understanding of codebases and writing truly usable code.
Replit has an incredibly robust prompt that has a variety of functionalities. Easily edit and search code at scale, install custom programming languages, and run arbitrary shell commands—all in one secure, collaborative environment. Cursor’s prompt stands out for its instructions on tool usage: “Only call tools when needed, and never mention tool names to the user — just describe what you’re doing.” This level of specificity is important for task completion without being overly burdensome to the user.
As Menezes points out, “every single company has an entirely different system prompt for the same base model.” This wide variation is a result of how public sector organizations have tailored their prompts to serve very specific operational requirements. He claims that these custom instructions allow businesses to squeeze maximum performance for particular domains and tasks. They’re really trying to tune the model to do exactly what’s needed for a particular domain, particular tasks,” he says.
In his pursuit to create the best AI-based solutions, Menezes has been internally “dogfooding” Superblocks’ products. Like the ubiquitous conversational AI agents his team has built, which help automate and simplify multiple lines of business. One agent uses Customer Relationship Management (CRM) data to identify new leads, and another monitors technical support metrics. To do so, they’ve developed an agent that assigns the appropriate tasks among their human sales engineers more effectively.
Menezes agrees that knowledge and mastery of system prompts is key to innovation potential. “This is basically a way for us to build the tools and not buy the tools,” he states. By developing their own AI models, companies can produce customized solutions. They achieve this by doubling down on highly specific prompting to fit their unique demands.
Superblocks recently closed a $23 million Series A funding round. This investment will enable them to develop safe enterprise applications in the new wild west of AI-based coding. The funding will help the company to explore more deeply the potential of system prompts. Most importantly, it will empower them to create exciting new tech solutions.
