Set max_tokens limits

Cap runaway output

The max_tokens parameter caps how many output tokens Claude can generate. Without it, Claude might write a 3,000-token essay when you only needed 200 tokens.

Sensible defaults

Set this intentionally based on what you actually need. For structured outputs (JSON, classifications), 256–512 is usually plenty. For long-form writing, 1024–2048.

Combine max_tokens with explicit instructions: *"Respond in under 3 sentences."* The parameter is a hard cap, the instruction shapes Claude's intent.