- 0
Stage 0 - Complete novice: make something appear
Goal: experience the magic and beat the blank page. Build a throwaway toy in a browser-based builder and feel the prompt-to-app loop.
Stage 0 playbook
Stage 0 - Complete novice: make something appear
Goal: experience the magic and beat the blank page. Build a throwaway toy in a browser-based builder and feel the prompt-to-app loop.
- Pick one browser-based builder (Lovable, v0, Bolt, or Replit) so there is nothing to install.
- Describe one tiny app in a sentence - a tip calculator, a personal landing page, a to-do list.
- Iterate by talking: "make the button blue", "add a dark mode". Watch how phrasing changes the result.
- Do not connect a real database, payments, or your own users yet. This stage is purely for reps.
Avoid this
Trying to build your "real" startup idea on day one. Build five disposable toys first; intuition for prompting is the actual skill.
- 1
Stage 1 - Advanced beginner: understand what it built
Goal: stop being a passenger. Learn to read the generated code well enough to spot when the AI is wrong.
Stage 1 playbook
Stage 1 - Advanced beginner: understand what it built
Goal: stop being a passenger. Learn to read the generated code well enough to spot when the AI is wrong.
- Ask the AI to explain its own code: "what does this file do, line by line?"
- Learn the shape of a project: frontend vs backend, where data is stored, where secrets live.
- Practice describing bugs precisely. "It's broken" gets nothing; "the form submits but the row never appears in the database" gets a fix.
- Start using version control (git/GitHub) so you can undo, and so a reviewer can watch your commits.
Avoid this
Accepting every suggestion blindly. The AI is a confident junior developer - fast, tireless, and occasionally very wrong.
- 2
Stage 2 - Building for real: graduate to agentic tools
Goal: ship something with actual users. Move to a code-aware editor or agent and start handling auth, data, and money.
Stage 2 playbook
Stage 2 - Building for real: graduate to agentic tools
Goal: ship something with actual users. Move to a code-aware editor or agent and start handling auth, data, and money.
- Move to a tool with full code access (Cursor, or an agent like Claude Code) so you control the whole project.
- Implement authentication and database access rules carefully - this is where AI-built apps leak data.
- Keep every secret in environment variables, never in the code or the repo.
- Review what an agent changes before you accept it. Autonomy is a force multiplier on mistakes too.
Avoid this
Letting an agent run unattended against production. Recursive loops and bad commands can run up huge bills or wipe data fast.
- 3
Stage 3 - Shipping: make it safe, scalable, and affordable
Goal: survive real users. Put a review pass between "it works" and "it's live" that covers the things demos never test.
Stage 3 playbook
Stage 3 - Shipping: make it safe, scalable, and affordable
Goal: survive real users. Put a review pass between "it works" and "it's live" that covers the things demos never test.
- Run a security pass: access control, input validation, exposed endpoints, and hallucinated/typosquatted dependencies.
- Run a scalability pass: N+1 queries, unbounded queries, full-table scans, and blocking I/O that fall over under load.
- Run a cost pass: estimate per-user and per-request cost before a flat price turns your best users into your biggest losses.
- Automate all three on every commit so the checks happen continuously, not once in a pre-launch panic.
Avoid this
Treating launch as the finish line. The vulnerability count in a codebase keeps climbing as you add features; the review has to be continuous.
Ready to ship what you build?
PeakStack reviews AI-built code for security, scalability, and cost on every commit - with the exact file, line, and fix.
Explore PeakStack