Prompting DeepSeek Well

Structure Beats Cleverness

Goal, context, constraints, format, example. Reasoning models use structure to plan; give them the plan''s raw materials.

Zero-Shot First

DeepSeek''s reasoning often needs no examples: ask directly, then add one example only if the format drifts. Fewer tokens, faster answers.

Steer the Trace

Ask for verification passes (check by substitution), alternative approaches, and confidence levels. The thinking trace is debuggable: quote the step where it went wrong and ask for a fix.