The agent built the feature in 20 minutes. It compiles, tests pass. Two weeks later you’re refactoring the whole thing because the design was wrong from the start.
When AI handles the typing, you can move faster — and faster means you can afford more experiments. More experiments means more learning. Experienced developers are leaning into this: treat the first pass as a spike — throwaway code whose only purpose is to answer a question about the problem.
Let the agent explore the solution space fast. Don’t agonise over structure. The goal isn’t production code, it’s discovery: Where do the design decisions cluster? What are the real trade-offs? What did you actually need?
Why? AI picks whatever is locally plausible. It doesn’t know you’re optimising for maintainability over speed, or that three teams will share this module. One engineer’s post-mortem after months of AI-heavy work: “I spent weeks following AI down dead ends. In hindsight, it might have been faster just thinking it through.”
Then start fresh. Anthropic’s own best-practices documentation says it plainly: “A clean session with a better prompt almost always outperforms a long session with accumulated corrections.” The spike gave you the understanding; the second pass gets your architecture.
đź’ˇ Next time you start something non-trivial: timebox a spike, extract what you learned, then git stash and build for real in a fresh session.
đź”— Lalit Maganti — “Eight years of wanting, three months of building with AI”
Do you ever intentionally throw away AI-generated code?
🟢 Yes — spike then rebuild is my normal flow
🟡 Sometimes, when I realise the structure was wrong
🔴 No — I refactor what the AI gives me
⚪ Hadn’t considered this approach
What’s the biggest thing you’ve thrown away and rebuilt?