Brad Menezes, CEO of enterprise vibe coding startup Superblocks, believes the following crop of billion-dollar startup concepts are hiding in nearly plain sight: the system prompts utilized by present unicorn AI startups.
System prompts are the prolonged prompts — over 5,000-6,000 phrases — that AI startups use to instruct the foundational fashions from corporations like OpenAI or Anthropic on the way to generate their application-level AI merchandise. They’re, in Menezes view, like a grasp class in immediate engineering.
“Each single firm has a totally totally different system immediate for a similar [foundational] mannequin,” he advised TechCrunch. “They’re making an attempt to get the mannequin to do precisely what’s required for a selected area, particular duties.”
System prompts aren’t precisely hidden. Clients can ask many AI instruments to share theirs. However they aren’t at all times publicly out there.
In order a part of his personal startup’s new product announcement of an enterprise coding AI agent named Clark, Superblocks supplied to share a file of 19 system prompts from among the hottest AI coding merchandise like Windsurf, Manus, Cursor, Lovable and Bolt.
Menezes’s tweet went viral, seen by nearly 2 million together with massive names within the Valley like Sam Blond, previously of Founders Fund and Brex, and Aaron Levie, a Superblocks investor. Superblocks announced final week that it raised a $23 million Collection A, bringing its complete to $60 million for its vibe coding instruments geared to non-developers at enterprises.
So we requested Menezes to stroll us by the way to research different’s system prompts to glean insights.
“I’d say the largest studying for us constructing Clark and studying by the system prompts is that the system immediate itself is possibly 20% of the key sauce,” Menezes defined. This immediate provides the LLM the baseline of what to do.
The opposite 80% is “immediate enrichment” he mentioned, which is the infrastructure a startup builds across the calls to the LLM. That half contains directions it attaches to a person’s immediate, and actions taken when returning the response, comparable to checking for accuracy.
Roles, context and instruments
He mentioned there are three components of system prompts to review: function prompting, contextual prompting, and gear use.
The very first thing to note is that, whereas system prompts are written in pure language, they’re exceptionally particular. “You mainly have to talk as if you happen to would to a human co-worker,” Menezes mentioned. “And the directions need to be good.”
Function prompting helps the LLMs be constant, giving each goal and persona. As an illustration, Devin’s begins with, “You’re Devin, a software program engineer utilizing an actual laptop working system. You’re a actual code-wiz: few programmers are as gifted as you at understanding codebases, writing practical and clear code, and iterating in your adjustments till they’re appropriate.”
Contextual prompting provides the fashions the context to think about earlier than performing. It ought to present guardrails that may, for example, cut back prices and guarantee readability on duties.
Cursor’s instructs, “Solely name instruments when wanted, and by no means point out device names to the person — simply describe what you’re doing. … don’t present code until requested. … Learn related file content material earlier than enhancing and repair clear errors, however don’t guess or loop fixes greater than thrice.”
Device use permits agentic duties as a result of it instructs the fashions the way to transcend simply producing textual content. Replit’s, for example, is lengthy and describes enhancing and looking out code, putting in languages, organising and querying PostgreSQL databases, executing shell instructions and extra.
Learning others’ system prompts helped Menezes see what different vibe coders emphasised. Instruments like Loveable, V0, and Bolt “concentrate on quick iteration,” he mentioned, whereas “Manus, Devin, OpenAI Codex, and Replit” assist customers create full-stack functions however “the output continues to be uncooked code.”
Menezes noticed a chance to let non-programmers write apps, if his startup might deal with extra, comparable to safety and entry to enterprise information sources like Salesforce.
Whereas he’s not but working the multi-billion startup of his desires, Superblock has landed some notable corporations as clients, it mentioned, together with Instacart and Paypaya International.
Menezes can be dogfooding the product internally. His software program engineers usually are not allowed to jot down inner instruments; they will solely construct the product. So his enterprise of us have constructed brokers for all their wants, like one which makes use of CRM information to determine leads, one which tracks assist metrics, one other that steadiness the assignments of the human gross sales engineers.
“That is mainly a manner for us to construct the instruments and never purchase the instruments,” he sais.