Common pitfalls
A few traps catch almost everyone in their first month:
- Accepting the first thing that runs. "It works" and "it's correct" are different claims. The first means the happy path didn't crash; the second means you checked the edges. Don't confuse them.
- Mega-prompts. Asking for an entire feature in one shot gives the model too much room to drift and gives you too much to review at once. Decompose. Ship the smallest reviewable slice, then build on it.
- Letting the context rot. As a grows, the model starts contradicting earlier decisions and re-introducing bugs you already fixed. When that starts, summarize the current state and restart fresh rather than fighting a confused thread.
- Trusting confidence as a signal. The model's tone is identical whether it's right or hallucinating. Its certainty tells you nothing. Your verification tells you everything.
- Outsourcing the decision, not just the typing. It's fine to ask "what are my options for X?" It's a mistake to let the answer pick your architecture for you without your judgment in the loop.