Sonnet 4.5 can code autonomously for 30 hours straight. But who wants to review 30 hours of AI-generated code?
Conference talks will show you vibe coding—Lovable, Cursor, Claude Code, Codex. Let AI think. Let it be creative. Ship fast. The problem: you lose control. The code looks like AI wrote it. Good luck taking over mid-feature with your bg-indigo-500 nextjs app.
I do raw dog coding. It works. The difference is policy-driven development.
I spent 12 months rebuilding official documentation for major frameworks. Created dependency graphs for every aspect of full-stack development. Defined step-by-step processes. I have been burning through ~$100/day in tokens.
The result: AI executes, doesn’t think. The code looks like you would write it. At any moment, you can pick up exactly where AI stopped and continue manually. Because it’s following your plan, your style, your tech stack - not improvising its own.
The payoff: you can build things that were never worth the investment. That internal tool sitting in your backlog for 3 years? Ship it finally now. That refactor everyone avoids? Execute it systematically.
What to ask when a supplier uses AI to write code
If you want a supplier to deliver faster with Claude Code or another coding agent, use the discussion to clarify responsibility:
- What working result will you accept, and how will it be checked?
- Who reviews generated changes and owns defects?
- Can your team understand, operate and change the delivered system without the supplier’s agent setup?
The tool name does not answer those questions. Reviewing the result, documenting decisions and agreeing a handover makes the proposed delivery approach assessable.
For an already defined need, see custom software development. If you are still deciding which backlog item or internal bottleneck is worth addressing, a company audit is a way to scope that decision before commissioning implementation.