AI Prompts Best Practices That Still Work in 2026

Updated October 26, 2026

The best practice for AI prompts is to write them the way you would brief a smart new hire who knows nothing about your situation: say exactly what you want done, give the context and the reason behind it, show one or more examples of good output, and state the format and length you expect. Short, vague prompts get generic answers because the model fills every gap with the most average guess. Most of the improvement comes from adding what you already know and forgot to say.

Some older habits have stopped working with the 2026 models: typing instructions in capitals, writing "think step by step" and prefilling the start of the answer. The sections below cover what to keep, what to drop and how to repair a prompt that is going wrong.

What should a good AI prompt include?

A good AI prompt includes the task, the context, the audience, the output format and, where the format or tone matters, an example. A role helps too. Anthropic's guide notes that even one sentence of role in the system prompt changes the model's focus and tone.

Here is the same request written two ways.

Weak:

Write a product description for our coffee grinder.

Strong:

You're writing for the product page of a small coffee equipment shop. Our customers are home espresso people who already own a decent machine and know what burr size means. Write a product description for the Aldo 64 grinder: 64mm flat burrs, stepless adjustment, single-dose hopper, $480. Lead with what it does for espresso consistency, keep it under 120 words, no exclamation marks, and end with one sentence on who should pick a cheaper grinder instead. Here's a description we like from our site, for tone: [paste].

The second prompt is longer, and it gets you something close to publishable on the first try because nothing is left to guess. The order matters less than people think for a prompt this short. What matters is that the information is there.

If you are not sure your prompt says enough, use the test in Anthropic's prompting guide: show it to a colleague with no background on the task. If they would be confused, the model will be too.

How do you write clear instructions for an AI model?

Clear instructions for an AI model are specific about the action, give the reason behind each rule and use normal wording, without shouting. Four habits do most of the work.

Say whether you want action or advice. Current models follow instructions literally. Anthropic gives the example that "Can you suggest some changes" may get you a list of suggestions, while "Change this function to improve its performance" gets the edits made. If you want the rewrite, ask for the rewrite.

Give the reason. "Never use ellipses" is a bare rule. "Your response will be read aloud by a text-to-speech engine, so never use ellipses" works better, according to Anthropic, because the model can generalize from the reason to cases you didn't list. When a rule matters, say why it exists.

Say what to do instead of what not to do. "Do not use markdown" works worse than "Write in smoothly flowing prose paragraphs." A related trick from the same guide: write your prompt in the style you want back. A prompt full of bullet points and headers tends to get bullet points and headers in return.

Drop the capitals. "CRITICAL: You MUST use this tool when..." was common advice a couple of years ago. Anthropic now says newer models are more responsive to the system prompt and that this kind of emphasis makes them overtrigger, so write "Use this tool when..." instead. Google's Gemini 3 guidance makes the same point from the other side: avoid overly persuasive language and be precise and direct.

One more: define any word that could mean two things. "Short" means 50 words to one person and two pages to another. "Formal" depends on the industry. Google's guidance for Gemini 3 specifically recommends defining ambiguous terms.

How many examples should you put in an AI prompt?

Put three to five examples in an AI prompt when format, tone or structure matters. That is Anthropic's recommendation, and the guide calls examples "one of the most reliable ways" to steer output. Google's Gemini prompting guide goes further and recommends always including few-shot examples, saying prompts without them are likely less effective.

How you pick the examples matters more than how many you give.

• Make them relevant to the actual task, not a loosely similar one.
• Make them varied. If every example opens with a question, every answer will too.
• Keep their formatting identical: same tags, same spacing, same separators.
• Wrap them so the model knows where they start and stop, for instance in <example> tags inside an <examples> block.

Google warns that too many examples can make the model overfit, copying surface details such as length, phrasing or topic instead of learning the pattern. If outputs start echoing one example word for word, cut that example or add a different one next to it.

For casual chat use, one good example beats a paragraph describing the tone you want. Paste a past email you liked and say "match this voice."

How should you structure a long AI prompt with documents?

Structure a long AI prompt by putting the documents first, the question and instructions last, and wrapping each part in labeled tags. Anthropic's guide applies this to inputs of 20,000 tokens or more, and Anthropic reports that putting the query at the end improved response quality by up to 30 percent in its tests. Google gives the same advice for Gemini 3: supply the context first and put the question at the very end, with a bridge such as "Based on the information above..."

Tags help the model tell instructions apart from material. A layout that works:

<documents>
  <document>
    <source>Q3 board deck</source>
    <content>...</content>
  </document>
  <document>
    <source>Customer interview notes</source>
    <content>...</content>
  </document>
</documents>

<instructions>
Based on the documents above, list the three risks the board deck
doesn't mention but the interviews raise. First quote the passages
you're relying on, then give your list.
</instructions>

The tag names don't have to be anything special. They should be descriptive and used consistently, and nested when the content has a hierarchy. Pick one system, tags or Markdown headings, and stick with it for the whole prompt; mixing them is the mistake Google's guidance warns against.

The "quote first" line is worth copying. Asking the model to pull the relevant passages before answering keeps it tied to your documents instead of to what it vaguely remembers about the topic. For answers that must come only from supplied material, Google's docs also provide a "strictly grounded" system clause you can adapt.

One placement detail differs. Google suggests putting critical constraints, persona and output format in the system instruction or at the very start, with the question at the end. So rules go at the top, material in the middle and the question at the bottom.

Should you still tell AI to think step by step?

You usually don't need to tell current AI models to think step by step, because the major ones now reason internally before answering. Google says Gemini 2.5 and 3 generate their own thinking text automatically. On Anthropic's newest models thinking is always on, and on several of them a prompt asking the model to write its reasoning out in <thinking> tags may be declined.

What works now is a general nudge. Anthropic recommends instructions like "think thoroughly" over a hand-written step-by-step plan, because the model's own reasoning often beats the plan a person would prescribe. Google suggests "Think very hard before answering" for heavy reasoning, noting that it costs extra thinking tokens. Save it for problems that need it, such as a tricky calculation, a bug hunt or a decision with many tradeoffs.

Asking for a self-check still pays off. "Before you finish, verify your answer against [your criteria]" catches errors reliably, especially in code and math, per Anthropic. The exception is Claude Opus 5, which verifies on its own; carrying the instruction over can make it check too much. That is the general pattern with model-specific tricks: Anthropic says to re-test any technique against your own evaluations before moving it to a different model.

How do you fix an AI prompt that gives bad results?

Fix an AI prompt by changing one thing at a time and comparing the outputs, starting with what is missing rather than what is wrong. Google describes prompt engineering as iterative and calls its own templates starting points. When the answer misses, work through this order:

1. Read the prompt as if you knew nothing. Most bad answers trace back to context that lived in your head.
2. Add or swap an example. This fixes format and tone problems faster than more description.
3. Rephrase the request. Google lists this first among its iteration strategies.
4. Recast it as an analogous task. Google's example: turn an open classification into a multiple-choice question.
5. Reorder the parts, moving examples, context and input around.
6. Split it. If one prompt asks for research, analysis and a polished draft, run three prompts and feed each output into the next. Google calls this chaining.

Length problems have their own causes. Models now have different defaults. Anthropic says Claude Opus 5 writes longer user-facing answers by default and that effort settings don't reliably shorten them, so ask for brevity in words ("under 150 words, no preamble"). Google says Gemini 3 answers directly and efficiently by default, so you have to ask outright if you want something conversational or detailed. If your prompt was tuned on one model, expect to adjust length instructions on another.

Dates catch people out too. A model only knows the world up to its training cutoff. For Gemini 3 Flash, Google suggests putting the current year and the knowledge cutoff (January 2025) in the system instructions for time-sensitive questions. The same idea works anywhere: tell the model today's date and paste in anything recent it needs.

Do AI prompt frameworks, generators and templates help?

AI prompt frameworks, generators and templates help as checklists, and they hurt when you fill them in mechanically. Most acronym frameworks reduce to the same parts covered above: who the model is, what to do, what it needs to know and what the output should look like. If an acronym reminds you to add the audience, use it. If it has you writing "Role: expert marketer" and nothing else, it is giving you a false sense of having briefed the model.

Prompt generators, the tools that expand a short request into a long structured one, are useful for seeing what a thorough prompt looks like. Their weakness is that they can't add the facts only you know: your customer, your constraints, the example you liked. Read the generated prompt and replace the placeholders with real details before you use it. The same goes for downloadable prompt PDFs and free courses. The underlying advice is the same everywhere; what you learn from practice is where your own prompts leave gaps.

Templates are worth keeping for tasks you repeat. Save the prompts that worked, with their examples, and reuse them. That beats any list of "100 best prompts."

Creative work is where generic advice fails most. Asking any model for "ideas" tends to produce the obvious ones first. A few things help: ask for 20 options instead of 5 and keep the last 10, ban the three ideas you've already considered, or ask for options from an unrelated field ("how would a theme park designer approach our onboarding?"). For the brainstorming stage specifically, StonedGPT is built to give unconventional angles and connect unrelated ideas, then you can take the strongest ones back to a more conventional model to sharpen. The method behind deliberately getting off the obvious path is covered in how to think outside the box.

What changed for developers using the APIs?

Developers using the Claude and Gemini APIs need to remove some settings that used to be standard. On the Anthropic side, prefilling the assistant's response on the last turn is no longer supported on Claude 4.6 and later and returns a 400 error, so use instructions to control format and preamble instead. Thinking now runs through thinking: {type: "adaptive"} with an effort parameter, and on Claude 4.7 and later setting budget_tokens returns a 400 error.

On the Google side, Gemini 4 and Gemini 3.6+ ignore custom temperature, top_p and top_k values and return deprecation warnings, with 400 errors planned for a future release, according to Google's prompting strategies page. For earlier Gemini 3 models, Google advises leaving those parameters at their defaults to avoid degraded reasoning. For complex JSON, use the structured output feature rather than describing the schema in the prompt.

For agent workflows, both companies push in the same direction. Google suggests steering how the agent plans (how far to break the problem down, how exhaustive to be) and how it executes (persistence, recovery from errors, risk assessment). Anthropic notes that Claude Fable 5.1 gives fewer progress updates during agentic work, so ask for updates if you want them.