Stop Letting LLMs Write Your Code. Do This Instead.
Handing your codebase to AI destroys your skills and ruins your software. Use LLMs for planning and research. Write the code yourself.
Tools · Source: Hacker News
What happened
Developers are burning out by handing their codebases over to AI agents. The promise of massive productivity gains is a trap. When you let large language models generate everything, your codebase turns into an alien wasteland. You lose the joy of programming. Your skills quickly atrophy because you stop practicing the core mechanics of your craft. After just a few weeks of full automation, returning to manual coding feels impossibly hard.
A Haskell developer recently outlined a survival guide for the AI era. The core rule is simple. Refuse the standard workflow of planning first and letting the agent code. Make the agents group around your preferred way to work. Plan together with the AI, but write the actual code yourself. Let the agents handle the incidental work that drains your energy.
This means using LLMs as advanced bookkeeping and research tools. Let them parse complex domain discussions into actionable tasks. Have them map out all the places in your codebase that need editing. Ask them to flag potential pitfalls before you start typing. When the prep work is done, you take the keyboard. You maintain full context of your system. You never have to rewrite generated slop.
Key facts
- Souls in the Great Machine — A book by Sean McMullen used to describe the dystopian reality of developers becoming mere cogs in a giant calculating unit.
- Generative Adversarial Networks — An AI architecture involving a generator and a discriminator, which the author suggests mimicking to create automated LLM review cycles.
- Haskell — The programming language the author uses, highlighting that developers who genuinely enjoy their tools have the most to lose from AI generation.
Why it matters
If you stop writing code, you stop being a good programmer. LLMs are much worse at producing readable, maintainable code than the hype suggests. They excel at writing code that only they can read. When a completely generated file inevitably contains a bug, you will lack the context to fix it. The landscape of your own product becomes alien to you. You are degraded from an active builder to a mere cog in a machine. You become entirely dependent on tech giants who can throttle your tokens at any moment.
Flipping this dynamic keeps you in control. Using AI for research, planning, and automated code review makes you faster without sacrificing quality. You avoid embarrassing technical debt because you understand the domain better than the agent does. More importantly, you retain ownership of your product. By keeping the complex logic in your own head, you reduce your reliance on expensive frontier models. You can easily swap them for smaller, open-source alternatives that run locally. You stop being a victim of token limits and start building on your own terms.
For builders
Build an automated review cycle
Never accept unreviewed AI output. Treat LLM coding like a Generative Adversarial Network. Use one agent to generate plans or minor code, and a separate reviewer agent to find logical holes and bugs. Do not look at the output until the reviewer agent gives it a passing grade.
Delegate research but verify sources
Do not let agents research a topic and blindly accept their conclusions as facts. Research in parallel using a standard search engine. Force the AI to cite specific sources for its proposals. This ensures you understand the domain well enough to make the final architectural decisions.
Restrict agents to low risk tasks
Only use coding agents for cleanup, routine work, and low-risk refactoring. If you have three interesting cases and seven boring ones, let the AI write the boring ones. Do not let them design complex systems from scratch. You pay the price when they hallucinate bad architecture.
My take
Vibe coding is a fast track to technical bankruptcy. If you outsource the core act of building to an opaque slot machine, you are not a founder anymore. You are just a middleman waiting for a tech giant to throttle your tokens. Write your own code.
Original reporting: Hacker News. This is my rewrite and opinion.