{"slug": "20-approaches-to-writing-better-ai-prompts", "title": "20 approaches to writing better AI prompts", "summary": "InfoWorld published a list of 20 prompt styles and structures for getting better results from large language models, including step-by-step instructions, few-shot examples, Q&A-style questioning, and prompts that ask the model to interrogate the user. The article notes that some teams employ dedicated prompt engineers who are sometimes more schooled in creative writing than computer science, and that instructions can at times run longer than the model's answer.", "body_md": "If AIs are supposed to be such magical time savers, why do we spend so much time and effort writing prompts? In some cases, the prompts can be longer than the answer!\n\nBut while there is some academic value to such philosophical navel gazing about the nature of prompting, the reality is that the most focused teams are devoting plenty of effort to creating just the right set of words to set the [large language model](https://www.infoworld.com/article/2335213/large-language-models-the-foundations-of-generative-ai.html) (LLM) in motion. They know that the best combination of input tokens can tickle the mystical pathways in the LLM’s weights as each token nudges the system into the best state that will produce the best answer. Some teams even have dedicated prompt engineers who are sometimes more schooled in creative writing than computer science.\n\nCreating good prompts is still an evolving art. Developers continue to experiment with different combinations of words and different sentence structures. Sometimes a different style or rhetorical stance leads to surprises.\n\nIn the interest of advancing the art, here is a list of 20 different prompt styles and structures for everyone to use and experiment with. Use them by themselves or even combine a few for a more effective hybrid strategy.\n\nA concise but detailed description of what the LLM is supposed to do, step after step, is one of the best approaches for getting consistent results. If you spell out the length, tone, and structure, the LLM or agent is usually able to generate a result that matches what you want, at least when compared with the instructions that are in the prompt. For some examples, the instructions can be longer than the answer—a statistic that leaves clever users scratching their heads and wondering if the AI is actually saving time.\n\nLLMs are mimics and they love to parrot bits from their training set or even your prompt. Giving the LLM a few examples is one of the simplest ways to get it to deliver what we need. Many models are graded on their ability to “follow instructions” with some of the test metrics designed to track how well they respond to direct examples in their prompt. This is especially useful when we need the data in a particular format like JSON or CSV.\n\nIn some cases, examples aren’t necessary or feasible. Asking a model to, say, translate a paragraph into French or Russian isn’t a situation where examples make sense. If the model is up to the task, the various words from the other language should be part of the training set. In these cases, the user can say little because the hard work has been done when the model was built.\n\nInstead of treating the LLM as a minion, the Q&A approach treats it as a sage with a deep ability to answer questions. A series of short questions followed by direct answers can often triangulate on the answer, delivering knowledge to the user. This only works, though, when the LLM is a tool for finding knowledge and, perhaps, teaching. It doesn’t work as well when the AI is meant to generate a block of writing. Some people feel that an interview or interrogation built from a series of short, direct questions reduces hallucinations, but that hasn’t been my experience. The LLMs are engineered to please us and they hallucinate when they’re grasping for an answer. More questions means more answers and more opportunities to hallucinate.\n\nLLMs are famously glib. Some users don’t want the glib answer, so they ask the LLM to respond not with an answer, but with a series of questions about the user’s task or goal instead. In other words, the model is asked to interrogate the human as a test. This approach is particularly popular for models that are deployed in schools, but it is helpful in any context where the user wants to gain understanding or force the model to think more deeply about a problem.\n\nWhat would Albert Einstein say? Or, maybe, Sylvia Plath? Or maybe just a helpful but not overly greedy plumber? Some prompts ask the LLM to assume the personality of a particular person, character, or type and answer as they would. This can be helpful is many situations when someone’s pose or personality is as important as the raw facts.\n\nWhen the LLM will be assuming the role of some historical figure, writing the prompt in the style of the desired time can help ensure accuracy. Adopting the jargon and argot of the era nudges the LLM into the style of documents in the training set from those years.\n\nSometimes the simplest approach is to write the first half of the answer and let the LLM fill in the rest. Think “The primary structural difference between DNA and RNA is that…”. This fits the natural architecture of the LLM, which is essentially a statistics-driven token assembly line.\n\nSome prompts ask the model to enter a loop of creation, evaluation, and revision. That is, you ask the model to work through several drafts before responding. Some agentic systems have this approach baked into their structure. Others can be prompted to work through several iterations.\n\nThe chain-of-thought prompt sets out the steps for the LLM or agent with precision so the machine will approach the job properly. First consider the axioms and then spell out every logical step from beginning to end. This is particularly important when relying upon [Model Context Protocol](https://www.infoworld.com/article/4029634/what-is-model-context-protocol-how-mcp-bridges-ai-and-external-services.html) (MCP) servers and tools, because a good chain-of-thought prompt can point out useful resources like databases or even some humans to include in the loop.\n\nSome problems naturally lend themselves to decision trees instead of linear chains. The tree-of-thought prompt tells the LLM to consider options, evaluate multiple options, weigh pros and cons, and then make decisions as the model works through a more complicated knowledge graph.\n\nWhen you want a long-form answer, it often pays to ask the agent to write a high-level outline before writing the final response. This is similar to the critique-and-revise loop but with an additional emphasis on thinking strategically about the overall structure. Some call this “[directional stimulus prompting](https://arxiv.org/abs/2302.11520)” because the high-level outline creates “subprompts” that use particular keywords or phrases as a stimulus to guide the model to the desired outcome.\n\nIt doesn’t hurt to let the LLM know what you really want. Adding a few extra words, phrases, or sentences can guide the model closer to your target. For example, instead of asking for a summary, ask for a summary and tell the LLM to focus on, say, the early years of the person’s life. A pointer or two can make a big difference with many models.\n\nSometimes it helps to ask the LLM to generate a list of facts or relevant details before creating the final document. Creating a good list simplifies the process for the LLM and lets it first concentrate on collecting data. Then the model can evaluate the information. In some cases, you might explicitly ask the agent to use one model for assembling the potential details and a second to analyze it. For some tasks, like coding, two models often work better than one.\n\nMany models react to emotional pleas like, “please answer this correctly because my job depends upon me finding an answer.” Benchmarks have [demonstrated the power of emotional pleas](https://intuitionlabs.ai/articles/llm-performance-high-stakes-emotional-prompts) with LLMs such as GPT, Claude, and Llama, something that illustrates just how much the models have learned from humans.\n\nSometimes it’s easier to tell the LLM what not to do. You may want the output to avoid long or obscure words. You may want to avoid jargon, clichés, or bullet points. LLMs are often very good at evaluating many potential responses, so eliminating some of them is a feasible way to produce a useful answer.\n\nIn some cases, LLMs can’t deliver certain answers and the prompts can spell out what topics or types of solutions the LLM should avoid. These legal limitations are a version of the negative prompts. In many cases, the system prompt can include a long list of types of questions that must be avoided to protect the company running the LLM.\n\nSome developers have found that including the right words in their prompt can overcome the restrictions spelled out in the system prompt or burned into the model during training. Some of the simplest approaches include claims like some child is about to die unless the model breaks a rule. Developing models that resist these jailbreaking efforts continues to be an active form of research for the companies that are selling inference services.\n\nWhen LLMs are asked to gather data for other purposes, getting the output in a particular format like [JSON](https://www.infoworld.com/article/2255837/what-is-json-a-better-format-for-data-exchange.html) or [Markdown](https://www.infoworld.com/article/3983394/what-is-markdown-lightweight-text-formatting-for-human-beings.html) makes it possible to feed the results into more traditional computational tools like databases. The simplest way to prompt an LLM is to give it a template and ask it to fill in details.\n\nSometimes we find ourselves at a loss for words. LLMs can often write good prompts themselves and it’s not uncommon for prompt engineers to ask the model for a first draft. In other words, prompting the LLM to write the prompt. When this first draft is fed directly to a second round, the process starts becoming very meta.", "url": "https://wpnews.pro/news/20-approaches-to-writing-better-ai-prompts", "canonical_source": "https://www.infoworld.com/article/4223596/20-approaches-to-writing-better-ai-prompts.html", "published_at": "2026-09-21 09:00:00+00:00", "updated_at": "2026-09-21 09:24:23.703175+00:00", "lang": "en", "topics": ["large-language-models", "generative-ai", "ai-tools", "natural-language-processing"], "entities": ["InfoWorld", "Albert Einstein", "Sylvia Plath"], "alternates": {"html": "https://wpnews.pro/news/20-approaches-to-writing-better-ai-prompts", "markdown": "https://wpnews.pro/news/20-approaches-to-writing-better-ai-prompts.md", "text": "https://wpnews.pro/news/20-approaches-to-writing-better-ai-prompts.txt", "jsonld": "https://wpnews.pro/news/20-approaches-to-writing-better-ai-prompts.jsonld"}}