Superblocks CEO: How to find a unicorn idea by studying AI system prompts

Brad Menezes, enterprise Vibe coding General Manager SuitableThe next product of the billion dollar starting statements is almost hiding in a straight landscape: the system is used by Unicorn AI beginnings.

System offers are long offers – more than 5000-6000 words – AI beginnings use companies such as AIRAI or anthropical companies to teach their application to create AI products. They are in the image as Master Klass, in Menezes.

“Every company has a completely different system offer [foundational] The model said, “TechCrunch said.” A special domain is trying to get the model to make exactly what is needed for special tasks. “

System offers are not fully hidden. Customers may want to share many AI means. But they are not always open.

Thus, the new product of its beginning, as a new product announcement of an enterprise with AI agent of Clark, offered Superblocks 19 Sharing a file of system instruction Some of the most popular AI coding products like Windsurf, Manus, Cursor, Favorite and Bolt.

Meneses Tweet went viralEarlier, the founders Foundation and Brex, Brech and Brex and Brex, Aaron Levie and Harun Levie, Aaron Levie, Aaron Levie’s, Aaron Levie’s, looked almost 2 million, including a superblocks. Suitable declared Last week raised a $ 23 million series, only reaches $ 60 million Vibe is directed to developers in enterprises for coding tools.

Thus, we wanted to walk us by learning how to learn the Menezes prepares another system.

“The biggest learning that the system has prepared and reads the system, the system itself is perhaps 20% of the secret sauce,” said menesians. This request gives the main part of what LLM do.

The other 80% said the “quick enrichment”, which is a starting infrastructure around the calls on LLM. This part includes actions and actions taken while returning the answer as a user’s desire and accuracy.

Rolls, context and tools

Said there are three parts for the study of three parts of the system: the role is asked, the use of contextual requests and tools.

The first thing to pay attention is that the system request is very specific, even if it is written in the natural language. “You should mainly talk to a man’s colleague as they want,” said Menezes. “And the hints should be perfect.”

The role you want to consistent LLMs and helps to give both goals and identities. For example, Devin’s is a software engineer who uses a real computer operating system. You are a real code-wiz: Several programmers are talented until you understand several programmers, functional and clean code, while writing and clean. “

Contextual desire gives the context to review before moving to models. For example, it must reduce costs and provide caravants that can clarify tasks.

Cereal’s instructions, “Never celebrate only call tools and tool names to the user – just describe what you do.

The use of the tool provides agent tasks, because the models just instruct the text to go beyond the Creator. The demonstration, for example, for a long time, explains shell commands and more to edit and correct the search code, installing language installation, placement and postgresql database.

Learning the system of others, helped menzes emphasized other vibe coding. Tools like your favorite, v0 and bolt “Pay attention to the fast iteration, “said,” Manus, Devin, Openai Codex and “users create full stacking applications, but” exit is still a raw code. ”

Menezes allow you to write applications without programs, such as security and enterprise information sources, such as security and enterprise information sources.

While still continued to begin a billion billion in the beginning of their dreams, Superblock, some prominent companies, Instacart and Paypaya global, this also lowered some outstanding companies.

Menezes are also a dog in the product. Its software engineers are not allowed to write internal means; They can only build the product. Thus, his business people, agents for all the needs, use CRM data to identify heads and follow support measurements that balancing the tasks of human sales engineers.

“This is a way to build tools and not receiving tools,” he said.