prompt-engineering

Use the golden pattern: state your goal in one sentence, provide context (tech stack and files), describe current behavior, specify desired behavior, list constraints on what not to change, and define acceptance criteria for verifying success.

Use gpt-4o-mini or claude-haiku for simple edits, gpt-4.1 or claude-sonnet for daily coding, gpt-5.2 or claude-opus for complex features, and o1/o3 reasoning models for architecture and hard debugging problems.

Paste the actual current code instead of describing it, be hyper-specific about function names and file paths, ask the AI to list its assumptions before coding, and explicitly state what must not be changed.

Use it when initial direct instructions fail or for complex logic. Ask the AI to explain its approach, assumptions, and edge cases before writing code, which surfaces mistakes early and improves solution quality.

Inline edit handles small single-file changes like fixing a typo, agent mode coordinates multi-file features like adding auth across five files, and composer suits complex reasoning tasks like refactoring an entire API architecture.

Start with cheaper models and upgrade only if quality suffers, place repeated context at the start to leverage prompt caching, chain large tasks into smaller prompts, and use structured outputs to avoid parsing retries.