Brad Menezes, CEO of Superblocks, champions a surprising idea: billion-dollar startup concepts might be tucked away in the detailed system prompts of today’s leading AI unicorns. These extensive prompts—often spanning over 5,000 to 6,000 words—guide foundational models from companies like OpenAI and Anthropic, helping them deliver AI applications that really click with user needs. Menezes calls this a master class in prompt engineering.
Every company crafts its own system prompt for a shared foundational model, customising it to suit unique tasks and domains. While the prompts aren’t exactly top secret—customers can request to see them—they rarely make their way into the public eye. In a bold move, Superblocks released a file containing 19 prompts from popular AI coding tools, including Windsurf, Manus, Cursor, Lovable, and Bolt, to mark the debut of their new coding agent, Clark. This release drew nearly 2 million viewers, including key players from Silicon Valley.
Superblocks recently secured $23 million in a Series A round, adding to a total of $60 million dedicated to developing vibe coding tools aimed at non-developers in enterprise settings. Menezes explains that while the system prompt is important—comprising about 20% of what makes these solutions work—the real magic lies in the surrounding “prompt enrichment.” This encompasses everything from the infrastructure built around large language model (LLM) calls to clear user instructions and robust response validation.
Menezes breaks down system prompts into three key elements: role prompting, contextual prompting, and tool usage. Role prompting sets the model’s personality and purpose, such as casting it as an expert software engineer. Contextual prompting ensures that the model sticks within defined parameters, which can help cut costs and clarify its tasks. Tool usage then equips the model to handle more complex tasks beyond just generating text—demonstrated by how Replit provides extensive instructions.
By analysing these prompts, Menezes observed distinct trends among vibe coders. Tools like Lovable, V0, and Bolt focus on rapid iteration, while others, including Manus, Devin, OpenAI Codex, and Replit, aim to build complete applications—even if the initial output is raw code. This careful study even sparked Menezes’s vision of enabling non-programmers to build apps, provided his startup can manage the security challenges and data access issues typical of enterprise systems like Salesforce.
Although Superblocks isn’t yet a billion-dollar enterprise, it boasts an impressive client roster that includes names like Instacart and Paypaya Global. Within the company, software engineers concentrate on developing the product while business teams work on creating the necessary tools. As Menezes puts it, they’re in the business of building their own tools rather than buying them.
