Why your AI coding projects are failing.
1. Monoliths. The AI will often default to monolithic design standards and principles. There are valid reasons to design and build monolithic apps, even complex ones. But they're not really "normal" anymore for most real software projects. The AI needs to understand this.
2. No definition of good. People expect the machine to read their minds. This is unreasonable. At the same time, the definition of "good" changes between developers, managers, projects, stacks, libraries, the level of the stack you're working at, and what industry you're in. Be explicit about what "good" means to you.
3. No planning. "Vibing" out a feature on an existing codebase in a tool like Claude Code is literally enough to burn your house down. Tell the machine to create documentation for everything. Proposals, plans, epics, breakdowns, QA, and coding standards. You'll be glad you did.
4. No process. Tasks need to be deterministic, and testable. This is a rule for building code in general -- at least if what you're doing is functional, rather than poetry.
5. No unit tests. You might hate unit tests as a developer, and not understand why you need them. In the case of AI augmented coding, you need unit tests for the same reason you've ALWAYS needed unit tests: to test your assumptions, and the mind model you're building the code with. Don't second guess this. They'll become your best debugging friend.
6. No security! Lacking rate limits, path traversal, and injection aren't accounted for when AI's generate code. Tell it to think about security best practices, and make a recommendation based on the codebase.
7. No QA loop. You haven't told your AI to peer review or QA the work. You've also had multiple lossy summaries, and you've probably got mocks and placeholder values that you don't find without knowing they're there.
8. Huge commits! You need to stay on top of this one, because when code moves a mile a minute.
9. Gigantic testing surfaces. You shouldn't want to build an app that isn't testable. But, seeing as you are human, you're not going to be able to effectively test 150,000 lines of code in one sitting, by yourself. Be sensible, and plan testing into your pipeline.
Amplification is the point
And there you have it.
All of the normal rules you would use for developing an app, but for AI!
Does all of this take the "easy" out of your AI automated sessions? Yes. Absolutely. Easy shouldn't be the point.
The point is amplification.