uniform, skewed, bimodal, anything 2. **n >= 30 is usually enough** (more for heavily skewed data) 3. **Standard error = sigma / sqrt(n)** — larger samples give more precise means