Modern reasoning models like DeepSeek-R1 and Claude 3.5 Sonnet perform drastically better when permitted to think out loud before emitting their final answer.
The XML Scratchpad Directive
markdown
Structure your response in two phases: <thinking> 1. Break down the user problem into independent sub-components. 2. Verify potential edge cases and counter-examples. 3. Draft intermediate calculations or pseudocode. </thinking> <solution> Provide the clean, polished final answer without referencing the thinking scratchpad. </solution>
This cleanly separates draft reasoning from your user-facing output.