Prompt engineering is often the simplest way to improve an LLM feature before building a more complex system.
2
Chain of Thought prompting breaks difficult tasks into intermediate reasoning steps, while few-shot prompting teaches a model through examples.
3
Reasoning models need a different approach: give them a clear task, encourage more reasoning when needed, and use few-shot examples sparingly.
Summary
Dan argues that prompt engineering remains a practical part of building LLM products. The prompt is only one part of the system, but it is often the easiest place to improve output. Before reaching for RAG or another complex architecture, teams should spend time testing whether a better prompt can solve the task. He explains Chain of Thought prompting, which asks a model to reason through intermediate steps, and few-shot prompting, which provides examples for the model to follow. Meta prompting uses an LLM to create or improve prompts, with the best approach varying between model providers. Reasoning models change these rules. Research around OpenAI o1, DeepSeek-R1, and Microsoft's Medprompt framework found that examples can reduce performance, while allowing more reasoning can improve results. Dan recommends starting with a clear task description and minimal prompting, then adding encouragement to reason only when the model needs more performance.
Prompt engineering is often the simplest fix for an LLM feature
Dan says prompt engineering can look unnecessary until someone has shipped an AI-based feature. Understanding what the model should do is difficult, and prompt design is the most accessible way to improve outputs. The prompt is also part of a larger product system that includes the surrounding architecture. Since everyone may have access to the same models, the way a product uses its prompts and architecture can create an advantage. He recommends spending time testing a prompt before building a complex RAG system. If the task can be solved with a prompt, the result is simpler to manage.
Chain of Thought prompting asks the model to work through intermediate steps
Chain of Thought prompting tells a model to reason about a problem before producing its answer. Dan describes it as breaking a problem into subproblems, which can also make troubleshooting easier. A basic zero-shot version adds an instruction such as "think step by step" or asks the model to take a breath before answering. Few-shot Chain of Thought gives the model an example problem together with the reasoning steps used to solve it. Dan also describes Automatic Chain of Thought and PromptHub's AutoReason template, which generates reasoning chains for a task.
Few-shot prompting teaches a model through examples rather than instructions alone
Few-shot prompting includes examples of the behavior or output the model should reproduce. Dan's example gives the model a client's brief and related content, then asks it to create content for a new brief. This can communicate tone and style more directly than describing them. He says most of the benefit usually comes from one or two examples. Those examples should cover different kinds of inputs the model may receive. Adding more examples can eventually reduce performance, so builders should avoid assuming that a larger collection is always better.
An LLM can help create and improve the prompts used by another LLM
Meta prompting uses an LLM to create, refine, or improve a prompt. Dan mentions free tools from Anthropic and OpenAI, as well as PromptHub's prompt enhancer. PromptHub lets users select a model provider and applies a different meta prompt because a prompt that works well for OpenAI models may not work as well for Anthropic models. Its workflow also lets users iterate on prompts, run them, and provide feedback. Dan presents this as a practical use of LLMs during the prompt engineering process.
Reasoning models change the advice about examples and instructions
Dan says reasoning models behave differently and need different prompting strategies. Microsoft's Medprompt research found that adding examples to a prompt for OpenAI o1 reduced performance, and DeepSeek researchers observed a similar result while developing R1. OpenAI also warned that extra context could overcomplicate a task and confuse the model. For reasoning models, Dan recommends a minimal prompt with a clear task description. Few-shot prompting should usually be avoided, or started with only one example and possibly two. Explicit instructions about how to reason can also hurt because reasoning is already built into these models.
More reasoning can improve results when a reasoning model needs extra performance
Dan refers to research showing that asking a reasoning model to think for longer can produce better results. Microsoft's prompt research compared a quick-response instruction with a request for more thought and found better results with extended reasoning. DeepSeek saw the same pattern as it trained R1: the model's thought process became longer, and accuracy and performance increased. Dan recommends encouraging more reasoning when a reasoning model is close to the desired level but still needs improvement. This should come after starting with a simple, clear task description.
"Overall, when you're using reasoning models specifically, minimal prompting."07:55
Who should watch
You are building an LLM feature and are considering a complex RAG or agent setup before testing whether prompt changes can solve the task.
Your outputs need a specific tone, format, or reasoning pattern, and you want to decide between examples, Chain of Thought instructions, and meta prompting.
You are using a reasoning model such as OpenAI o1 or DeepSeek-R1 and have seen few-shot examples make the results worse.
