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

Brad Menzes, CEO of Enterprise VB Coding Startup SuperblockBelieve that the next crop of billion -dollar startup ideas is hidden in a simple sight: the system used by the existing Unicorn AI startups is prompts.

System requests are more than 5,000-6,000 words-which AI startups use Foundational models to instruct Foundational models on OpenAI or ethnographic companies how to create their application-level AI products. They are in the Menzes View like Prompt Engineering Master Class.

“Each single company has a completely different system prompt for the same [foundational] Model, “He told TechCrunch.” They are trying to get the model to do exactly what is needed for certain domains, specific tasks. “

The system prompts are not exactly hidden. Customers can ask a lot of AI tools to share them. However, they are not always universally available.

Thus an enterprise coding name of Clark is part of the new startup announcement of the AI ​​agent, as part of the announcement of superblock To share a file of 19 systems request From some of the most popular AI coding products such as Windsorph, Manus, Cursor, Bolt and Bolt.

Of the manizes Tweets have been viralValley, like Sam Blond, has been viewed by about 2 million with large names, the previous Foundation Fund and Brex and Super Blocks investors Aaron Levi. Superblock Declaration Last week it raised in a 23 million series, Its total is being brought in $ 60 million Its vibe is ready for non-sectors in the enterprise for coding equipment.

So we asked us to continue by requesting our insights insights on how we can study the system of another.

Menzes explained, “The biggest learning of our clerk and the prompt of the system is that the system’s prompt itself may be 20% of the secret sauce,” the Menzes explained. This prompt gives LLM the baseline of what to do.

The other 5% is “prompt enrichment” he said, it is an infrastructure that creates a startup LLM’s calls. In this section it includes testing for accuracy such as the steps taken when the instructions attached to a user’s prompt.

Introduction, context and equipment

He said that there are requests for studying three parts of the system: Introduction Prompting, relevant prompting and use of equipment.

The first thing to note is that the system’s requests are written in natural language, but they are exceptionally specific. “You have to say that you want a human colleague with a human colleague,” said Menzes. “And the instructions should be perfect.”

The role prompting helps LLMs to be consistent, giving both purpose and personality. For example, Devin began with, “You are a software engineer using a real computer operating system.

The relevant prompting models give the context to consider before acting. It should be supplied to maintenance which, for example, reduce the cost and ensure precision in the functions.

Cursor’s instructions, “Call the tools only if needed and never mention the tools to the user – just describe what you are doing …… don’t show the code until you ask.

The use of equipment enables agents because it instructs models how to go out of text simply. For example, the transcripts describe long and editing and search codes, language installed, postgressQL databases setup and searching, performing shell commands and more.

Studying other systems requests helps to see what other VIB coders emphasized. Loving, equipment like V0 and Bolts ““Focus on the quick repetition,” he said, where “Manus, Devin, OpenA Codex and Reflit” helps users to create full-destination applications but “the output is still a raw code.”

Menzes saw the opportunity to let non-programmmers write the opportunity to write, if his startup could further manage such security and access to enterprise data sources such as safety and salesforce.

Although he is not yet running a multi-billion startup of his dreams, Superblock has landed some significant companies as customers, saying it, including Inschart and Pepaya Global.

Menzes are internally dogging the product. His software engineers are not allowed to write internal equipment; They could simply make products. So its business people have created agents for all their needs, such as one that uses to detect leads using CRM data, one that supports metric, the other that balances human sales engineers’ assignments.

“This is basically a way to make the tools for us and not buy the equipment,” he is Sais.