ONLINEAGENT_OPS 2026.Q3 HOME HUBSMODELSBUILDINGCOMPANIESFILMMAKINGSECURITYLIVING WITH ITSKILLSALL HUBS, ONE TERMINAL → ARTICLES CRAFT RECORD BLOG TOOLS SKILLS SEARCH /
HOME›ARTICLES›PROMPTING TECHNIQUES EXPLAINED
ARTICLES · GUIDEINTERMEDIATE

Prompting Techniques Explained: Zero-Shot, Few-Shot and More

The main prompting techniques in plain English: zero-shot, few-shot, role, context, format, chain of thought and chaining, with when each helps.

READ14 min
WORDS2,785
SECTIONS14
TYPEGUIDE
CHECKED12 OCT 26

Prompting techniques are repeatable ways of writing a request to an AI so the answer comes back closer to what you need, such as zero-shot (just ask), few-shot (show examples), role and audience, and chain of thought (ask for the steps).

The main ones are zero-shot (just ask), few-shot (show examples), role and audience, context, format, labelled sections, chain of thought (ask for the steps), and prompt chaining (split the job into stages). Most everyday prompts need only the first few; the rest help with harder or repeated jobs. Each one below comes with what it is, when it helps, a short example and what the AI companies themselves say about it.

TL;DR — THE SHORT VERSION
  • Start zero-shot: a clear request with no examples. Add the others only when the answer misses.
  • Wrong format or tone? Show two or three varied examples (few-shot).
  • Wrong level or voice? Say who it is writing as and who will read it (role and audience). A role sets tone; it does not add skill.
  • Long or messy input? Label the parts, and put long documents first with the question last.
  • Hard reasoning? On older or “fast” models, ask for the steps. On reasoning models, just state the goal.
  • Big job? Chain it: outline, check, then write. Or draft, review, improve.

The techniques at a glance

The plain point: each technique fixes a different kind of miss. Pick the one that matches what went wrong.

TechniqueWhat it isReach for it when
Zero-shotIs: instructions only, no examples.Use it for: clear, common tasks. Always try it first.
Few-shotIs: instructions plus a few worked examples.Use it for: a format, label set or tone that is hard to describe.
Role and audienceIs: who the AI writes as, and who reads it.Use it for: getting the voice and the level right.
Context and reasonIs: the background and why you need it.Use it for: answers that are generic or miss your situation.
FormatIs: the shape of the answer: list, table, word limit, JSON.Use it for: output you will paste somewhere or skim.
Labelled sectionsIs: headings or tags that separate instructions from material.Use it for: long prompts with pasted documents or data.
Chain of thoughtIs: asking for the steps before the answer.Use it for: maths and logic on models that do not reason on their own.
Prompt chainingIs: one job split into several prompts in a row.Use it for: jobs with stages you want to check.
Self-checkIs: asking it to review its own draft against your criteria.Use it for: important drafts, before you do your own check.

The techniques stack. A good everyday prompt is often zero-shot with a role, some context and a format, which is the shape what makes a good prompt calls its six elements. This page explains each one on its own so you know which lever to pull.

Zero-shot prompting: just ask

The plain point: describe the task clearly and let the AI use what it already knows. No examples.

Google’s Gemini guide defines the two basic modes in one breath: prompts with a few examples are few-shot, and prompts with none are zero-shot.1 IBM calls zero-shot “often the default” strategy for a new problem.2

ZERO-SHOT — ADAPT THE BRACKETS
Summarise the article below in [three] bullet points for [a busy manager]. Each bullet under [20] words. End with one line on what we should decide. [paste the article]

When it is enough: most of the time, for everyday writing, summarising and explaining. Full page: zero-shot prompting.

Few-shot prompting: show, do not describe

The plain point: paste a few examples of input and the answer you want, then give it the new input.

OpenAI says examples steer a model toward a new task “rather than fine-tuning the model”, and asks for “a diverse range of possible inputs with the desired outputs”.3 Anthropic suggests three to five examples that are relevant and varied.4 Google adds that the examples should all share the same structure and formatting.5 One example is called one-shot.

FEW-SHOT — ADAPT THE EXAMPLES
Rewrite each product note as a friendly one-line description for our website. Note: oak board, 40cm, oiled Description: A solid oak chopping board, oiled and ready to use. Note: linen tea towel, pack of 2, washed Description: Two soft, pre-washed linen tea towels. Note: [your next note] Description:

Watch out: the AI copies your examples closely, including their mistakes and their length. On OpenAI’s reasoning models, the company says to try without examples first.6 Full page: few-shot prompting.

TAKEAWAY

If you have rewritten your description of the format three times, stop. Two or three examples will usually say it faster.

Role and audience: who is writing, who is reading

The plain point: one sentence on the voice to use and one on the reader changes the tone and the level of detail.

All three big vendors build this in. Anthropic says a role “focuses Claude’s behavior and tone”, and that even one sentence makes a difference.7 OpenAI’s guide opens its suggested instructions with an “Identity” section describing the assistant’s purpose and communication style.8 Google’s prompting guide for Gemini calls it the persona.9

ROLE AND AUDIENCE — ADAPT THE BRACKETS
You are [a school nurse writing to parents]. Explain [what to do if a child has head lice] for [busy parents with no medical background]. Calm, practical, no jargon, under [150] words.

Watch out: “You are a world-class expert” does not make the model know more. The vendors describe a role as setting behaviour and tone, so put the real specifics of the task next to it. Prompting myths covers this one. A role you want in every chat belongs in a system prompt or custom instructions.

Context and the reason why

The plain point: the AI only knows what you tell it. Give the situation, and say why you need the answer.

Google advises including the information the model needs “instead of assuming that the model has all of the required information.”5 Anthropic goes a step further: explaining why an instruction matters helps the model give a more targeted answer.7

CONTEXT AND REASON — ADAPT THE BRACKETS
I am [a volunteer running a small food bank]. We [lost our Tuesday driver and need a replacement within two weeks]. Write a short appeal for [our community Facebook group]. It will be read on phones, so keep it to [5] short lines. The goal is [three people messaging us to ask about the role].

When it helps: whenever an answer is correct but generic. The missing piece is usually the situation, not a cleverer instruction.

Format and structured output

The plain point: say what shape you want back. A table, three options, a word limit, plain paragraphs.

Google treats the response format as an instruction in its own right.5 Anthropic’s tip is to say what to do rather than what not to do: ask for flowing paragraphs instead of banning bullet points.7 For developers, there is a stricter version: OpenAI’s API can return data in a fixed JSON layout, a feature it calls Structured Outputs.8 In a chat app, asking for a table with named columns gets you most of the way.

FORMAT — ADAPT THE BRACKETS
Compare [three broadband deals I paste below] in a table with these columns: provider, contract length, speed, setup fee, what happens at the end of the contract. Use only the details I paste. Write “not stated” where something is missing. [paste the deal details]

Labelled sections: keep instructions apart from material

The plain point: when a prompt mixes your instructions with pasted text, label each part so the AI does not confuse them.

OpenAI suggests Markdown headings and XML tags to mark where each part of a prompt begins and ends.8 Anthropic says tags like <instructions> and <context> reduce misreading, and that long documents work best near the top, above your question.7 Google gives the same order for long material: context first, the question at the end.5

LABELLED SECTIONS — ADAPT THE CONTENT
<document> [paste the tenancy agreement] </document> <instructions> Using only the document above, list every date by which I must do something, with the clause number. If a deadline is unclear, say so instead of guessing. </instructions>

Plain capital headings (DOCUMENT:, TASK:) work too. The point is the separation, not the exact symbols.

Chain of thought: ask for the steps

The plain point: asking the AI to work through a problem before answering can help on maths and logic, but newer reasoning models already do this on their own.

The idea comes from a 2022 paper by Wei and colleagues, which found that showing a model a few worked examples with their steps written out improved its reasoning.10 A few months later, Kojima and colleagues found that simply adding “Let’s think step by step”, with no examples at all, also helped.11

Then the models changed. OpenAI says that on its reasoning models, asking them to think step by step is unnecessary, and can sometimes hinder.12 Anthropic keeps the manual version as a fallback for when its model’s own thinking is switched off.7

CHAIN OF THOUGHT — FOR MODELS WITHOUT A THINKING MODE
A recipe for 4 people needs [300 g] of flour. I am cooking for [7]. Work it out step by step, show each step on its own line, then give the final amount on the last line.

Watch out: tidy steps are not proof. Anthropic’s researchers warn that the reasoning a model shows need not match how it actually reached the answer, so check the answer itself.13 Full page: chain of thought.

TAKEAWAY

Know which kind of model you are using. On a thinking model, state the goal and what a good answer looks like; on a fast model, asking for the steps can still help.

Prompt chaining: one job, several prompts

The plain point: split a big task into stages, and check each stage before the next one builds on it.

Google’s guide describes it plainly: make each step a prompt, and the output of one becomes the input of the next.5 Anthropic says newer models handle most multi-step reasoning internally, but chaining is still useful when you want to inspect what happens in between, and it names the most common chain: draft, review, refine.7

CHAIN STEP 1 — OUTLINE FIRST
Give me an outline only, no full text: a [one-page] guide for [new volunteers] on [our first-day routine]. Five headings, one line under each saying what it covers.
CHAIN STEP 2 — AFTER YOU HAVE CHECKED AND FIXED THE OUTLINE
Here is the corrected outline. Write the guide from it, under [400] words, in plain friendly English. Do not add any rule that is not in the outline. [paste your corrected outline]

When it helps: anything with natural stages: outline then draft, draft then translate, extract then summarise. Full page: prompt chaining.

Self-check: ask it to review its own work

The plain point: after a draft, ask the AI to check it against your criteria and list problems. Then check it yourself.

This is the review step in Anthropic’s draft, review, refine chain above. A related research idea, self-consistency, takes several separate reasoning attempts and keeps the answer they agree on most; its 2022 paper reported gains on maths and common-sense tests.14 In a chat app, the everyday version is simple: if an answer matters, regenerate it or ask in a fresh chat, and treat disagreement as a sign to check.

SELF-CHECK — ADAPT THE CRITERIA
Check the draft above against these rules: [under 200 words; every date matches my notes; no promises I did not make; plain English]. List each rule it breaks, quote the sentence, and suggest a fix. Do not rewrite the whole draft.

Watch out: a model reviewing itself can miss its own mistakes. Its review is a second pair of eyes, not the last one. The review loop covers checking in order.

Which technique, when

The plain point: diagnose the miss, then add one technique at a time.

  • The answer is generic: add context and the reason.
  • The facts are fine but the shape is wrong: add a format; if that fails, add examples.
  • The tone or level is off: add a role and name the reader.
  • It mixes up your instructions and your pasted text: label the sections and put the material first.
  • It gets reasoning wrong: switch on a thinking mode if your app has one; otherwise ask for the steps.
  • The job is big: chain it, and check between stages.

Change one thing at a time so you can see what helped. How to test a prompt sets out the method.

Do these techniques work the same on every AI?

The plain point: the basics travel; the details do not. Re-test when you switch models.

OpenAI says different types of model can need different prompting, and that results can differ even between versions of one model.8 The clearest example on this page: Google recommends always including examples, while OpenAI says its reasoning models often do better without them. Prompting for LLMs sets out what differs between ChatGPT, Claude and Gemini. Image and video tools follow their own rules; see what the vendors actually say.

Questions people ask

What is the difference between zero-shot and few-shot prompting?

Zero-shot gives the AI instructions only; few-shot adds a few examples of the input and the answer you want. Google says few-shot examples are often used to control the formatting, phrasing and scope of the answer.5

Is chain-of-thought prompting still useful?

On reasoning models, mostly not: OpenAI calls “think step by step” unnecessary there. On faster models, or with thinking switched off, Anthropic still suggests asking the model to think through the problem first.

What is the best prompting technique?

There is no single best one. The vendors’ shared advice is to be clear and specific and give the context the model needs; the other techniques are fixes for particular misses. OpenAI’s help page for ChatGPT puts clarity first.15

Are prompting techniques the same as prompt engineering?

They are its toolkit. Prompt engineering is the wider practice of designing and refining prompts; the techniques here are the moves it uses. For long-running AI work, the bigger job is now managing everything the model sees, covered in context engineering.

Where to go next

QUICK CHECK · 3 QUESTIONS

Pick an answer to see if you got it. Nothing is recorded.

1. Which technique does the page say to try first?

  • Prompt chaining, splitting every job into stages
  • Chain of thought, asking for every step
  • Zero-shot: a clear request with no examples
Show the answer

Answer: Zero-shot: a clear request with no examples. The page says start zero-shot and add the other techniques only when the answer misses.

2. The answer keeps coming back in the wrong format or tone. What does the page suggest?

  • Show two or three varied examples of what you want
  • Tell it that it is a world-class expert
  • Make the request as short as possible
Show the answer

Answer: Show two or three varied examples of what you want. The page says if you have rewritten your description of the format three times, two or three examples will usually say it faster.

3. Does telling the AI it is a world-class expert make it know more?

  • Yes, it unlocks hidden knowledge
  • No, a role sets tone and voice, not skill
  • Yes, but only on paid plans
Show the answer

Answer: No, a role sets tone and voice, not skill. The page says a role sets behaviour and tone, so put the real specifics of the task next to it.

SOURCES
  1. 1Google, Prompt design strategies, Gemini API docs, last updated 17 Sep 2026, read at source 12 Oct 2026“Prompts that contain a few examples are called few-shot prompts, while prompts that provide no examples are called zero-shot prompts.” ↑
  2. 2IBM, What is zero-shot prompting?, undated, read at source 12 Oct 2026“A zero-shot prompting strategy is often the default prompt engineering strategy when tasking a model with a new problem.” ↑
  3. 3OpenAI, Prompt engineering, API docs, undated, read at source 12 Oct 2026“Few-shot learning lets you steer a large language model toward a new task by including a handful of input/output examples in the prompt, rather than fine-tuning the model.” … “When providing examples, try to show a diverse range of possible inputs with the desired outputs.” ↑
  4. 4Anthropic, Prompting best practices, Claude developer docs, undated, read at source 12 Oct 2026“Include 3–5 examples for best results.” … “Diverse: Cover edge cases and vary enough that Claude doesn’t pick up unintended patterns.” ↑
  5. 5Google, Prompt design strategies, read at source 12 Oct 2026“Make sure that the structure and formatting of few-shot examples are the same to avoid responses with undesired formats.” ↑a“You can include instructions and information in a prompt that the model needs to solve a problem, instead of assuming that the model has all of the required information.” ↑b“You can give instructions that specify the format of the response.” ↑c“When providing large amounts of context (e.g., documents, code), supply all the context first. Place your specific instructions or questions at the very end of the prompt.” ↑d“Chain prompts: For complex tasks that involve multiple sequential steps, make each step a prompt and chain the prompts together in a sequence. In this sequential chain of prompts, the output of one prompt in the sequence becomes the input of the next prompt.” ↑e“Few-shot prompts are often used to regulate the formatting, phrasing, scoping, or general patterning of model responses.” ↑f
  6. 6OpenAI, Reasoning best practices, undated guide, read at source 12 Oct 2026“Reasoning models often don’t need few-shot examples to produce good results, so try to write prompts without examples first.” ↑
  7. 7Anthropic, Prompting best practices, read at source 12 Oct 2026“Setting a role in the system prompt focuses Claude’s behavior and tone for your use case. Even a single sentence makes a difference”. ↑a“Providing context or motivation behind your instructions, such as explaining to Claude why such behavior is important, can give Claude more information about your goals and help it deliver more targeted responses.” (wording re-read at source 12 Nov 2026) ↑b“Tell Claude what to do instead of what not to do”, with the example: instead of “Do not use markdown in your response”, try “Your response should be composed of smoothly flowing prose paragraphs.” ↑c“Wrapping each type of content in its own tag (for example, <instructions>, <context>, <input>) reduces misinterpretation.” … “Place your long documents and inputs near the top of your prompt, above your query, instructions, and examples.” ↑d“Manual chain-of-thought (CoT) prompting as a fallback. When thinking is off, you can still encourage step-by-step reasoning by asking Claude to think through the problem before it answers”. ↑e“Explicit prompt chaining (breaking a task into sequential API calls) is still useful when you need to inspect intermediate outputs or enforce a specific pipeline structure.” … “The most common chaining pattern is self-correction: generate a draft → have Claude review it against criteria → have Claude refine based on the review.” ↑f
  8. 8OpenAI, Prompt engineering, read at source 12 Oct 2026“Identity: Describe the purpose, communication style, and high-level goals of the assistant.” ↑a“In addition to plain text, you can also have the model return structured data in JSON format - this feature is called Structured Outputs.” ↑b“When writing developer and user messages, you can help the model understand logical boundaries of your prompt and context data using a combination of Markdown formatting and XML tags.” ↑c“Some prompt engineering techniques work with every model, like using message roles. But different model types (like reasoning versus GPT models) might need to be prompted differently to produce the best results. Even different snapshots of models within the same family could produce different results.” ↑d
  9. 9Google Workspace, Writing effective AI prompts (Prompting guide 101), undated, read at source 12 Oct 2026“The four main areas to consider when writing an effective prompt are: Persona, Task, Context, Format”. ↑
  10. 10Wei et al., Chain-of-Thought Prompting Elicits Reasoning in Large Language Models, arXiv:2201.11903, 28 Jan 2022, read at source 12 Oct 2026“generating a chain of thought -- a series of intermediate reasoning steps -- significantly improves the ability of large language models to perform complex reasoning.” ↑
  11. 11Kojima et al., Large Language Models are Zero-Shot Reasoners, arXiv:2205.11916, 24 May 2022, read at source 12 Oct 2026“we show that LLMs are decent zero-shot reasoners by simply adding ‘Let’s think step by step’ before each answer.” ↑
  12. 12OpenAI, Reasoning best practices, read at source 12 Oct 2026“Some prompt engineering techniques, like instructing the model to ‘think step by step,’ may not enhance performance (and can sometimes hinder it).” … “Since these models perform reasoning internally, prompting them to ‘think step by step’ or ‘explain your reasoning’ is unnecessary.” ↑
  13. 13Anthropic, Reasoning models don’t always say what they think, 3 Apr 2025, read at source 1 Oct 2026 (as cited on this site’s chain-of-thought page): “There’s no specific reason why the reported Chain-of-Thought must accurately reflect the true reasoning process”.↑
  14. 14Wang et al., Self-Consistency Improves Chain of Thought Reasoning in Language Models, arXiv:2203.11171, 21 Mar 2022, read at source 12 Oct 2026“It first samples a diverse set of reasoning paths instead of only taking the greedy one, and then selects the most consistent answer by marginalizing out the sampled reasoning paths.” ↑
  15. 15OpenAI Help Center, Prompt engineering best practices for ChatGPT, “Updated: 2 months ago” as shown, read at source 12 Oct 2026 in a browser“Ensure your prompts are clear, specific, and provide enough context for the model to understand what you are asking.” ↑

OpenAI, Prompt engineering and Reasoning best practices, API docs, read at source 12 Oct 2026; OpenAI Help Center, Prompt engineering best practices for ChatGPT, read at source 12 Oct 2026 in a browser.

Anthropic, Prompting best practices, read at source 12 Oct 2026; Anthropic, Reasoning models don’t always say what they think, 3 Apr 2025, read at source 1 Oct 2026.

Google Workspace, Writing effective AI prompts, read at source 12 Oct 2026. IBM, What is zero-shot prompting?, read at source 12 Oct 2026.

Wei et al., Chain-of-Thought Prompting Elicits Reasoning in Large Language Models, 2022; Kojima et al., Large Language Models are Zero-Shot Reasoners, 2022; Wang et al., Self-Consistency Improves Chain of Thought Reasoning in Language Models, 2022; all read at source 12 Oct 2026.

Older read dates are those recorded on this site’s linked detail pages. The table, the “which technique, when” list and all example prompts are this site’s own, not claims from these sources.

Free printable AI cheat sheets, and AI news every week.

Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.

SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK
aipromptgeneer://$ about this siteAI PROMPT GENEERAI explained in plain English. Every claim links to its source, with the date it was checked.NEWSLETTERWhat changed in AI each week, in plain English.SUBSCRIBE FREE ↗AI PROMPT GENEER · MACHINE ASSISTANCE DISCLOSEDEVERY ARTICLE CARRIES ITS OWN CHECKED DATE