Ok so I was honestly losing my mind trying to guide DeepSeek-R1 with dense prompts, but I learned my lesson fast! Adding "think step-by-step" actually messed with its native flow and caused errors. Honestly, my go-to is being super direct. Seriously, let its internal reasoning do the work! dont over-engineer prompts cuz it knows what its doing. Less is lowkey more with this beast!! gl!

Interested in this too

In my experience, you definately wanna avoid adding "think step-by-step" because DeepSeek-R1 is basically already hardwired for deep reasoning. If you force it, the output gets super redundant and honestly a bit messy. I’ve compared three main styles lately: Zero-shot with strict constraints, Few-shot with logic traces, and Markdown-structured system prompts. The "Zero-shot + Constraints" approach is lowkey the best choice. Instead of telling it *how* to think, just define exactly *what* the final answer must include—basically use an "Output Requirements" block at the end of your prompt. I found that if I use the DeepSeek-V3 API for simple boilerplate and then switch to the DeepSeek-R1 API for the logic-heavy bits, it works like a charm. Few-shotting is kinda hit or miss cuz if your examples dont match its internal reasoning style, it might actually glitch out. Stick to clear Markdown headers and let its native CoT do the heavy lifting. It’s way more efficient and keeps things concise... works well for me anyway! gl