Two people can ask the same AI the same question and walk away with very different answers. Usually, the difference isn’t the AI. It’s how they asked.
That’s what the Prompting group on the Periodic Table of AI is about. It’s the biggest group on the chart, with 20 red tiles, and it’s the one you can put to work today. You don’t need to code. You just need to type.

The last post covered how models get trained. This one is about what happens after, when you sit down and talk to one.
The basics
Prompt
A prompt is whatever you send to the AI. A question, a request, a pasted email with “make this friendlier.” It’s the starting point for everything else in this post.
The biggest single improvement most people can make is to say more. “Write a cover letter” gets you something generic. “Write a cover letter for a warehouse supervisor job, I have six years of forklift and inventory experience, keep it under 250 words” gets you something you might actually send.
System Prompt
A system prompt is a set of standing instructions that sits behind the conversation. You usually don’t see it. Companies use it to tell a chatbot what it’s for, how to behave and what to avoid.
Many chat apps now let you write your own version, often called custom instructions. If you always want short answers or plain English, putting it there means you don’t have to repeat yourself.
Role Prompting
Role prompting means telling the AI who to be. “You’re a patient math tutor.” “You’re an editor who hates filler words.” The role nudges the tone, the vocabulary and what the AI pays attention to.
It works best when the role is specific and fits the job. “You’re an expert” is vague. “You’re a high school chemistry teacher explaining this to a nervous freshman” gives the model something to work with.
Showing instead of telling
Zero-shot
A zero-shot prompt asks the AI to do something without giving it any examples. “Translate this sentence into Spanish.” Most of what people type every day is zero-shot, and modern models handle a lot of it well.
Few-shot
A few-shot prompt includes a handful of examples first. If you want product descriptions in a certain style, you paste in two or three you like, then ask for a new one.
This idea got famous with the 2020 paper that introduced GPT-3, titled “Language Models are Few-Shot Learners.” The researchers showed that a big enough model could pick up a new task from just a few examples typed into the prompt, with no extra training at all.
Helping the AI think
Chain of Thought
Chain of thought means getting the AI to work through a problem step by step before it answers. It helps most with math, logic and anything with several moving parts.
A 2022 paper from Google researchers put this on the map. They gave models a few examples that showed the reasoning steps, not just the answers, and the models got much better at math word problems and similar tasks. Later that year, another team found you could get a similar boost without any examples, just by adding “Let’s think step by step.” In their tests on one set of math problems, that single line raised a model’s accuracy from about 18% to about 79%.
Extended Thinking
Extended thinking is chain of thought built right into the model. Some newer models can spend extra time reasoning on their own before they reply. You don’t have to ask them to think step by step, because they already do.
It’s useful for hard problems. For a simple question, it can just mean a slower answer.
Self-consistency
Self-consistency means asking the same question several times and going with the answer that comes up most often. It works because the AI doesn’t take the same path every time. One run might slip up, but if most runs land on the same answer, that answer is more likely right.
It comes from a 2022 paper by researchers at Google, who used it on top of chain of thought.
Self-critique
Self-critique means asking the AI to check its own work. After it answers, you say something like “Review your answer for mistakes and fix anything that’s wrong.” It catches some errors, though not all of them. An AI can be confidently wrong about its own work, too.
Organizing your prompt
XML Tags
XML tags are labels in angle brackets that wrap parts of your prompt, like <instructions> and </instructions>. They help the AI tell your instructions apart from your examples and the material you pasted in.
Anthropic recommends them for its Claude models. Its documentation says there’s no special set of “best” tag names. Just pick names that make sense for what’s inside and use them the same way every time.
Structured Output
Structured output means asking for the answer in a fixed format, like a table, a numbered list or JSON, a format computers can read. It matters most when the answer is going into another program or a spreadsheet, where a chatty paragraph would break things.
Prefilling
Prefilling means writing the first few words of the AI’s answer yourself, so it continues from there. Start its reply with an opening bracket and it’s far more likely to answer in that format. It’s mostly a tool for developers, since most chat apps don’t let you do it.
Prompt Templates
A prompt template is a reusable prompt with blanks to fill in. “Write a product description for [product] aimed at [audience] in under [number] words.” Once you’ve found a prompt that works, a template lets you reuse it without starting from scratch.
Breaking big jobs into steps
Prompt Chaining
Prompt chaining means splitting a big task into smaller prompts, where each answer feeds into the next. First, outline the article. Then, write section one. Then, edit for tone. Each step is easier to get right, and easier to fix when it goes wrong.
Meta-prompting
Meta-prompting means using AI to help write your prompts. You describe what you’re trying to do and ask the AI to draft a prompt for it. It’s a handy way to get unstuck, and you’ll often see ideas you wouldn’t have thought to include.
Iterative Refinement
Iterative refinement is the habit of improving an answer through rounds of feedback. You rarely get the perfect answer on the first try. “Shorter.” “Less formal.” “Add an example about small businesses.” Each round gets you closer.
Feeding the AI the right information
Context Engineering
Context engineering is the bigger cousin of prompting. It’s about everything the AI can see when it answers, including instructions, documents, past messages and tool results, not just the words you typed. Since a model only knows what’s in its context window, choosing what goes in there matters a lot.
Grounding
Grounding means tying the AI’s answer to real sources you provide. You paste in the document and say “Answer only from this text.” It’s one of the best ways to cut down on hallucination, because the AI has something solid to work from instead of its memory.
Tuning and testing
Temperature
Temperature is a setting that controls how much randomness goes into the AI’s word choices. Low temperature gives you more predictable, focused answers, which suits facts and code. High temperature gives you more variety, which suits brainstorming and creative writing. Most chat apps set it for you, but developer tools usually let you change it.
Evals
Evals are tests you run on your own prompts. You collect a set of example questions with good answers, then check how your prompt does across all of them. If you change the prompt, you run the evals again to see whether it got better or worse.
It’s the difference between “this prompt seems good” and “this prompt gets 47 out of 50 right.”
How they fit together
You don’t need all 20 at once. Start with the basics: say what you want, give context, and show an example when the format matters. Reach for chain of thought when the problem is tricky, grounding when facts matter, and chaining when the job is big. Once you have a prompt you’ll reuse, turn it into a template and test it with evals.
Where to go next
The next group on the chart is Model Families, the big names like GPT, Claude, Gemini and Llama, and that’s the next post.
For now, try one thing from this post the next time you use an AI. If you only try one, make it Few-shot. Two good examples can change the answer more than a paragraph of instructions.
1 thought on “AI Prompting: 20 Ways to Get Better Answers From AI”
Comments are closed.