Brad Menezes, CEO of Enterprise Vibe Coding Startup Superblocks, believes that the next harvest of billion dollar start-up ideas is hidden in almost simple vision: the system requests used by existing startups of the unicorn AI start-ups.
System requests are the long requests over 5,000 to 6,000 words, with which AI startups instruct the basic models of companies such as Openaai or Anthropic to generate their AI products at the application level. They are in Menezes view like a master class in prompt engineering.
“Every single company has a completely different system request for the same (basic) model,” he told Techcrunch. “You try to do the model exactly what is required for a specific domain, certain tasks.”
System requests are not exactly hidden. Customers can ask many AI tools to share theirs. But they are not always publicly available.
As part of the new product termination of the startup of his own startup of a company -KI agent named Clark, Superblock was offered to share a file with 19 system requirements of some of the most popular AI coding products such as windsurf, manus, cursor, graciousness and bolts.
Menezes’ tweet was viral, viewed by almost 2 million, including big names in the valley such as Sam Blond, formerly Founders Fund and Brex, and Aaron Levie, a super blocks investor. Last week, Superblocks announced that it collected an extension of Serie A $ 23 million and that Total series A brought in to $ 60 million for its Vibe coding tools, which were geared towards non-developers at Enterprises.
So we asked Menezes to take other insights through studying system requests.
“I would say that the greatest learning for us building Clark and reading the system requests in the fact that the system request may be 20% of the secret sauce,” said Menezes. This prompt gives the LLM the basis of what to do.
The other 80% are “immediately enriched,” he said what the infrastructure is, which builds a startup around the calls to the LLM. This part contains instructions that are attached to the user’s command prompt, and actions when returning the answer, e.g. B. the review for accuracy.
Roles, context and tools
He said there are three parts of system requests for studying: roles, context request and tool use.
The first thing to notice is that system requests in the natural language while they are exceptionally specific. “Basically, they have to speak as if they would be with a human employee,” said Menezes. “And the instructions have to be perfect.”
Roll prompt helps the LLMS to be consistent and to give both the purpose and personality. For example, Devin’s begins: “You are Devin, a software engineer who uses a real computer operating system. They are a real code wiz: Few programmers are talented as they understand code bases, for writing functional and clean code and itteries in their changes until they are correct.”
The context request gives the models into account the context before action. It should provide guidelines that, for example, reduce the costs and ensure the clarity of the tasks.
Cursors instructions: “Call up tools only if necessary and never mention the user tool names – just describe what you do.
The use of tools enables agent tasks because the models are exceeded how to go beyond creating text. The replits, for example, is long and describes the editing and search code, the installation of languages, setting up and querying postgreSQL databases, execution of Shell commands and much more.
Studying the system requests of other Half Menezes to see what other Vibe coders emphasized. Tools such as lovable, V0 and bolts “Concentrate on a quick iteration, ”he said, while“ Manus, Devin, Openaai Codex and Replit ”help users create full stack applications, but“ the output is still raw code ”.
Menezes saw the opportunity to have non -programmers write apps if his startup could do more, e.g. B. Security and access to company data sources such as Salesforce.
While he has not yet operated the startup of his dreams of several billion years, Superblocks has landed some remarkable companies as customers, including Instacart and PayPaya globally.
Menezes also sets the product internally with dog food. His software engineers are not allowed to write internal tools. You can only build the product. So his business people have built up agents for all their needs, e.g.
“This is basically a way for us to create the tools and not buy the tools,” he says.
Correction: This story has been updated to clarify that it was recently announced that Round was an expansion round and the series to update a total amount.