Make AI Obey: A Detailed Prompt Engineering Guide to Boost Large‑Model Logic

DeepSeek reasoning model (R1)

DeepSeek provides two models. DeepSeek‑V3 is trained on ordinary question‑answer data using the instruction, input, output format and does not learn a reasoning chain, so its logical inference ability is weak. DeepSeek‑R1 is trained on a Chain‑of‑Thought (CoT) dataset that adds a think field containing step‑by‑step reasoning. This enables the model to generate both the answer and the intermediate reasoning steps, improving its inference capability.

The DeepSeek web UI allows users to select the "Deep Thinking (R1)" option, which routes queries to the reasoning model.

Few‑shot prompting

Beyond the four essential elements (instruction, context, input, output), adding a few examples to the prompt supplies richer context. In a translation task, providing a single example corrected ChatGPT’s output, demonstrating that even one illustrative sample can dramatically improve answer quality.

Chain‑of‑Thought (CoT) prompting

For more complex logical problems, few‑shot prompting may still fail. A math problem about counting apples ("I bought 10 apples, gave 2 to a neighbor and 2 to a repairman, then bought 5 more and ate 1") is answered incorrectly by standard prompting. Adding a CoT prompt that includes the reasoning steps (the think field) yields the correct answer. This approach follows the method described in Google’s 2022 paper Chain‑of‑Thought Prompting Elicits Reasoning in Large Language Models .

Zero‑shot Chain‑of‑Thought

The same CoT principle can be applied without explicit examples. Inserting a magic phrase such as "Let’s think step by step" before the query triggers the model to generate a reasoning chain on its own, leading to the correct answer for the apple‑counting problem.

Universal prompt workflow

The four techniques—DeepSeek‑R1, few‑shot prompting, CoT prompting, and zero‑shot CoT—can be combined into a hierarchical template. The recommended sequence is:

1. Query the reasoning model (repeat if needed, choose the most frequent answer)
2. If ineffective, add few‑shot examples (1‑5 samples)
3. If still ineffective, try zero‑shot chain‑of‑thought
4. If still ineffective, add explicit chain‑of‑thought steps

This step‑by‑step workflow helps solve increasingly complex reasoning tasks with large language models.