Superblocks CEO: How to find the idea of ​​unicorn by studying the artificial intelligence system

Brad Minizis, CEO of Enterprise Vib Coding Startup SuperblocksIt is believed that the next crop of billiards of dollars is almost hidden in sight: it calls for the system used by startups in the current unicorn AI.

System claims are long claims-more than 5000 to 6000 words-which are used by startups from artificial intelligence to direct constitutive models from companies such as Openai or Antarbur on how to create AI products at the application level. It is, at Menezes, like a major category in fast engineering.

“Every company has a completely different system for the same model.” “They are trying to get the model to do exactly what is required for a specific field and specific tasks.”

The system claims are not exactly hidden. Customers can ask many artificial intelligence tools to share their basics. But it is not always available to the public.

So, as part of the new product announcement to start its operating agent for the AI ​​agent coding the institution called Clarak, Superblocks offer To share a file of 19 system claims Some of the most popular artificial intelligence coding products such as Windsurf, Manus, Cursor, Loveable and Bolt.

Minizis A viral tweet went,, Which are seen by nearly two million people including the big names in the valley such as Sam Blond, formerly from Founders Fund and Brex, and Aaron Levie, Superblocks investor. Superblocks Declare Last week, the A series of $ 23 million collected, Its total is 60 million dollars For coding tools for non -developers in institutions.

So we asked meezes the Sirna through how to study the system of others calling for the assembly of visions.

“I would like to say that the biggest learning for us is to build Clark and read claims through the system is that the demand for the system itself is perhaps 20 % of the secret sauce,” Mainesis explained. This router gives LLM the foundation line of what to do.

The other 80 % said the “immediate enrichment”, which is the infrastructure around which a start -up company revolves around the call to LLM. This part includes guidelines that are linked to the user’s demand, and the actions taken when returning the response, such as verifying accuracy.

Roles, context and tools

He said that there are three parts of the system’s requests to study: the role of the role, the contextual student, and the use of tools.

The first thing to notice is that although system claims are written in the natural language, they are exceptionally defined. “You have to speak essentially as if you were going to get a fellow human work,” MINIESES said. “The instructions must be perfect.”

The role of the claim helps LLMS to be consistent, while giving both the purpose and personality. For example, Devin begins, “You are Devin, a software engineer who uses a real computer operating system. You are a real symbol: a few talented programmers as you are in understanding the bases of the code, writing a clean and clean code, and repeating your changes to correct.”

The contextual context gives models the context that must be taken into account before acting. Greens that can, for example, should save costs and ensure clarity in tasks.

Cursor tools, “Communication tools only when needed, and never mention the names of the tools for the user – just row what you do … Do not display a symbol unless it is presented … Read the content of the relevant file before editing and fix clear errors, but do not guess or specify the episode more than three times.”

The tool allows the tasks the agent because it guides models on how only the text exceeds. Replit’s, for example, long and describes the editing of software instructions, installation of languages, preparing postgresql databases, to implement Shell and more orders and more.

The study of others in Meezes helped to know what other programmers emphasized. Tools like Lovable, V0 and Bolt “He said: “Focus on fast repetition,” while helping users “Manus, Devin, Openai Codex and Refore” to create complete applications, but “the output is still a raw symbol.”

Menezes saw an opportunity to allow non -programmers in writing applications, if you can start operating more, such as safety and access to institution data such as Salesforce.

Although he has not yet managed the start of billions of dreams, Superblock got some prominent companies as customers, including Instacart and PayPaya Global.

Menezes is also bid farewell to the product internally. His software engineers are not allowed to write internal tools; They can only build the product. So people from his business built agents for all their needs, such as those that use CRM data to identify expected customers, which track support standards, and another balance between human sales tasks.

“This is essentially a way to build tools and not buy tools,” Sais said.