I let AI write 100% of my code for 30 days. Here’s what broke

Description and use

AI rapidly scaffolds and generates boilerplate, but fails to maintain system coherence, introduces architectural drift, and writes plausible‑but‑wrong code. Human review and a separate skeptic are essential for safe production.

Problem

Can an AI write all application code without human coding, and what failures arise when relying solely on AI for production code?

Outcome

AI can rapidly scaffold and generate boilerplate, but it struggles with system‑wide coherence, introduces subtle architectural drift, writes plausible‑but‑wrong code, and fails to handle edge cases. Human review and separate skeptic roles are essential to catch these issues.

Workflow

• 1. Set a rule: no manual application code for 30 days.
• 2. Prompt AI to scaffold Next.js app with Postgres, Drizzle, and auth.
• 3. Generate boilerplate: CRUD endpoints, Zod schemas, table components.
• 4. Review AI output line by line, rejecting or re‑prompting as needed.
• 5. Handle configuration, secrets, and dashboard actions manually.
• 6. Log any “breaks” (issues that required manual intervention).
• 7. After 30 days, ship the product to production.
• 8. Analyze the 9 logged breaks and adjust guardrails for future use.

Evidence

The AI is brilliant at the file in front of it and blind to the one three folders over.

Source

The billing webhook handler it wrote acknowledged Stripe events before persisting them.

Source

When something broke that the AI couldn't see, asking it to fix it produced changes, not fixes.

Source