Here I collect AI coding tips I find useful, based on my experience and this HN discussion. The most important one: write the damn code. Don’t become a prompt refiner.
Use project rules
Put stuff the model repeatedly gets wrong into CLAUDE.md. Document coding conventions, domain terms, how to run tests. Update it whenever the model annoys you the same way twice.
Plan before coding
Use plan mode. For big tasks, have the model generate a spec, then architecture docs, then TODO lists. Only then let it implement small, well-scoped tasks.
Give it ways to self-check
Provide test commands. Let it run tests in a loop until they pass. For UI, attach browser tools so it can see the actual rendered page.
Treat it like a new dev
Break work into small tasks. Give technical descriptions plus relevant files. Let it plan and ask questions. You stay focused on architecture; it does plumbing.
Start from a reference
Hand-code one instance well. Commit it. Tell the model to follow that pattern for the rest.
Use it where it shines
AI is great for: repetitive similar changes, JSON wrangling, generating tests for existing code. It’s worse for designing systems from scratch with vague requirements.
Reset often
Don’t use endless chats. One conversation per task. Instructions stop influencing output after many turns. Start fresh frequently.
Be explicit
Never just say “build feature X” and leave. Explain the desired end state. Make the model restate requirements. Review every diff.
Think specific tools, not “AI”
Ask: do I need better autocomplete? One-off code examples? Boilerplate? Use LLMs where you understand the domain well enough to verify results.