Studying AI system prompts can help founders spot how AI products define users, workflows, tools, and risks—but it cannot, by itself, reveal a proven unicorn opportunity. Superblocks CEO Brad Menezes’s argument is best treated as a way to generate business hypotheses: inspect what successful AI products are designed to do, then look for valuable workflow problems their prompts and surrounding systems do not solve.
What a system prompt can—and cannot—tell you
A system prompt is a set of higher-priority instructions that shapes an AI application’s behavior: its role, task, context, constraints, and sometimes how it uses tools. It is only one part of the system. A user prompt is the immediate request; retrieved context is information supplied at runtime; tool definitions specify available actions; and post-processing can validate, filter, or route the model’s response.
That distinction matters because a prompt visible in a public example may not be the complete production setup. Runtime context, retrieval, model routing, fine-tuning, tool permissions, safety checks, and human review may all be absent. Treat a prompt as evidence of product assumptions, not a complete blueprint or proof of how the product performs.
In a June 7, 2025, TechCrunch interview, Superblocks CEO Brad Menezes argued that comparing prompts can expose differences in how AI products approach similar jobs, even when they use similar foundation models. He described the system prompt as roughly 20% of an AI product’s “secret sauce,” with the other 80% in “prompt enrichment”—his estimate, not an independently measured industry statistic. TechCrunch’s interview with Menezes is the source for that thesis and the examples below.
Recommended Free Tools
#1 Best Overall
- TURN IDEAS INTO REALITY – Feeling stuck with your idea and not sure where to start? This guided journal helps you write a complete business plan so you can gain clarity and move forward with confidence as an entrepreneur.
- SIMPLE DAILY PRACTICE – 13 guided journaling sections with over 100+ business planning prompts. Make this business planner part of your routine to build momentum and work toward your business goals in just 5 minutes a day.
- BUSINESS PLANNER FOR ENTREPRENEURS – Use this guided journal to define your vision, understand your customers, evaluate competitors, plan expenses, and create a clear roadmap for launching your business.
- PERSONAL GROWTH – Designed as a personal growth workbook to help you reconnect with your purpose, prioritize well-being, and build a business plan centered around meaningful impact.
- PREMIUM ECO-FRIENDLY JOURNAL – Crafted with 100% FSC-certified recycled paper, a recycled cardboard cover, and wrapped in luxurious linen. This entrepreneur planner blends sustainability with thoughtful design.
What to look for when comparing prompts
Role: what is the product claiming the AI should be?
Look at whether the AI is framed as an assistant, reviewer, operator, or autonomous worker—and whether it is supposed to explain, recommend, act, or verify. The interview cited Devin’s prompt as describing a highly capable software engineer operating a real computer environment. Such framing signals positioning and expected initiative; it does not establish that the product reliably meets that standard.
For a founder, the business question is whether the role points to a distinct user need. “Help me write code,” “complete a software task,” and “operate an enterprise workflow under controls” imply different buyers, risk levels, and product requirements, even if the model underneath is similar.
Context: what does the AI need to know before acting?
Context instructions can reveal operational lessons. The interview described Cursor instructions that included inspecting relevant files before editing, using tools only when needed, and avoiding repeated speculative fixes. These are examples attributed to the interview, not a complete or current specification of Cursor.
When reviewing a prompt, note what the system is told to inspect, trust, avoid assuming, or do when information is missing. Instructions about retries, tool use, and error recovery may point to friction such as wasted time, unreliable changes, or unnecessary cost. They are clues to investigate—not proof of a market gap.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Tools: can the AI only answer, or can it change something?
Tool descriptions reveal the boundary between generating text and taking action. The interview described Replit’s prompt as covering code editing and search, language installation, PostgreSQL setup and queries, and shell commands. A tool-enabled agent may read files, query business systems, execute code, or make changes; each capability raises distinct questions about permission, reversibility, and oversight.
Map tools by what they can read, change, execute, or query. A product connected to a database or deployment system is supporting a workflow, not merely a chat. The business opportunity may lie in making that workflow safe and useful, rather than in writing a more elaborate instruction.
What Superblocks said it learned from 19 coding-product prompts
In connection with announcing Clark, its enterprise coding agent, Superblocks assembled a file of 19 prompts collected from or associated with popular AI coding products, including Windsurf, Manus, Cursor, Lovable, and Bolt, according to the TechCrunch interview. The account does not establish that the prompts were complete, current, or gathered under consistent conditions, so the collection should not be treated as a controlled market study.
Menezes characterized Lovable, v0, and Bolt as emphasizing fast iteration. He described Manus, Devin, OpenAI Codex, and Replit as helping users create full-stack applications while still producing output that was largely raw code. These are his interpretations, not independent benchmark findings or current product rankings.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →Rank #3
The opportunity he identified was enterprise application creation for non-programmers, with security and access to business data sources such as Salesforce in view. In practical terms, that suggests a gap between generating code and delivering an internal application people can safely use: connecting enterprise data, respecting access controls, and turning output into an operational tool. That is a synthesis of the reported strategy, not a claim that prompt analysis alone validated demand.
A repeatable way to turn prompt observations into startup hypotheses
1. Build a public, permissioned sample
Use material vendors have published or shared with permission: prompt examples, product documentation, developer guides, public demonstrations, open-source configurations, tool schemas, and API documentation. Do not extract hidden instructions, bypass access controls, or assume that a prompt is free to reuse simply because it can be viewed.
2. Record the same fields for each product
Normalize the observations so that you compare products on the same questions rather than on prompt length or writing style.
| Field | Questions to record |
|---|---|
| User and job | Who is meant to use the product, and what task is it supposed to complete? |
| Role and autonomy | Is the AI an adviser, agent, reviewer, or operator? Does it suggest, ask, or act? |
| Context | What information must it inspect or receive, and what must it not assume? |
| Tools and permissions | What can it read, change, execute, or query? Is approval required? |
| Guardrails and verification | What actions are restricted, and how does the system check its work? |
| Failure handling | How does it recover from tool errors, ambiguity, or a bad result? |
| Output | What format or level of explanation is expected? |
| Buyer and value | Who would pay, and what measurable result could justify the purchase? |
| Unmet need | What valuable part of the workflow appears to remain unresolved? |
3. Separate conventions from meaningful signals
Generic directions such as “be helpful” or “ask clarifying questions” may be common conventions, not evidence of a business opportunity. Repeated operational instructions can be more revealing: inspect source files before editing, validate changes, limit retries, avoid unnecessary tool calls, preserve state across steps, or work with external business systems. Even these are only signals. Verify the underlying user pain with direct observation and customer evidence.
Free tools Windows power users keep installed
One-click scans. No signup required.
4. Map what happens around the model
For a promising workflow, diagram the full sequence: what data is gathered before the model call, which actions are taken during execution, and what checks happen after generation. Ask who approves a consequential action, how quality is measured, what happens when the model is wrong, and how the system recovers.
This is where Menezes’s “prompt enrichment” idea becomes practical. The surrounding product may need retrieval, current customer data, permission checks, task-specific context, tool orchestration, state management, retries, testing, audit logs, or rollback. Those components can matter more to a customer than the visible wording of an instruction.
5. State a buyer-centered hypothesis
Translate the gap into a specific claim: “For [buyer] who loses [time, money, or confidence] because [workflow failure], build [system] combining [model capability] with [data, tools, controls, and validation].” If you cannot identify the buyer, the costly or frequent problem, and a measurable outcome, you have an interesting prompt observation—not yet a business case.
6. Test demand and reliability before scaling
- Observe the current workflow and ask prospective buyers how often the problem occurs, what it costs, and what they already pay to address it.
- Prototype the narrowest useful task, then define a task-specific evaluation set, failure categories, and escalation rules.
- Measure model and tool costs, human review, latency, and the value produced per completed workflow.
- Test real deployment requirements, including integrations, identity, permissions, auditability, and customer approval processes.
- Check whether customers will adopt and pay for the result—not merely praise a demonstration.
Why the valuable part is often outside the prompt
A prompt can reveal what a product wants an AI to do, but durable product value may depend on everything required to make that action dependable. Before a model call, the application may retrieve relevant records, identify the user’s permissions, select tools, and remove irrelevant or sensitive information. During execution, it may manage state, enforce budgets, require approval, or run code in a sandbox. After generation, it may validate a schema, run tests, screen for policy violations, log actions, request human review, or roll back a change.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Best Value
Enterprise use makes these surrounding requirements especially important. Access to sensitive data, permission boundaries, audit trails, reliability, and change management are not solved by making the system prompt more persuasive. An agent with powerful tools needs least-privilege access, monitoring, and a recovery path because tool access increases the consequences of failure.
Prompt wording alone is also easy to imitate. More substantial advantages may come from proprietary workflow data, deep integrations, evaluation sets, customer-specific configuration, compliance infrastructure, embedded distribution, or operational know-how. Whether any of those become defensible depends on execution and customer adoption; a prompt comparison cannot establish that in advance.
Common mistakes in prompt-based opportunity hunting
- Copying instructions instead of solving the workflow: Reproducing a prompt does not provide its data, integrations, validation, or customer access. Map the complete workflow and identify what the buyer cannot do today.
- Treating a vendor’s positioning as proof: Public demonstrations and product claims can be aspirational. Compare documentation, tool descriptions, observed behavior, and customer evidence, and attribute claims to their source.
- Equating prompt length with quality: A long prompt may contain redundant or legacy rules. Ask what behavior each instruction is intended to change and how you would measure that change.
- Ignoring the budget owner: A technically impressive agent may not have a buyer. Identify who feels the pain, controls the budget, and judges the outcome.
- Building a generic wrapper: A broad interface around a foundation model can be easy for an incumbent to replicate. Anchor the product in a narrow workflow, difficult integration, reliable operations, or distribution advantage.
- Skipping evaluation and recovery: A polished demo does not show performance on real cases. Define how errors are detected, escalated, and reversed before giving an agent consequential access.
- Using material without authorization: Public visibility does not automatically resolve confidentiality, contractual, copyright, or security concerns. Use public, permissioned, or voluntarily disclosed material.
Use prompt study alongside other discovery methods
Prompt analysis is one source of product hypotheses, not a substitute for learning how work actually gets done. Workflow observation can expose handoffs and workarounds that instructions do not mention. Customer interviews and support tickets can reveal repeated delays or failures. API and integration analysis can show where systems are difficult to connect; regulatory changes can create new documentation or audit needs; open-source issues can expose persistent user pain; and spend analysis can highlight expensive manual work.
These methods help answer questions prompts cannot: whether the pain is urgent, who will pay, how large the market is, whether the solution can be distributed, and whether customers will keep using it. Menezes’s thesis can help a founder notice what AI products assume. It cannot prove that the resulting idea will become a unicorn.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Fix the driver behind crashes, sound loss and screen glitches3Repair Windows errors before they cause bigger problemsQuick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
