Few-shot Prompting Examples: 10 Patterns that Actually Work
- Harold Bell
- 3 days ago
- 10 min read
TL;DR
|
Short Answer Few-shot prompting examples follow a consistent structural pattern: clear delimiters separating examples, three to five varied input-output pairs that span the realistic input space, consistent formatting across examples, and the actual task input presented in the same format. The most common production patterns include classification, structured extraction, content generation, format transformation, and style matching. The structural discipline matters more than the specific wording in any single example. |
I get asked for few-shot prompting examples constantly. Marketing teams know the technique works in theory but want concrete patterns to copy and adapt. This article is the practical reference — ten patterns I use across content production, classification, and structured output workflows at MQL Magnet, with the specific structures that produce reliable output.
Each example shows the full prompt structure. Treat them as starting points. The structural shape (delimiters, example count, formatting consistency) is more important than the specific wording. Adapt the wording for your task; keep the structure intact.
Pattern 1 binary classification
Use when you need the model to sort inputs into one of two categories. Common applications include qualifying inbound content, flagging potentially-relevant items in a backlog, and content moderation.
Prompt structure:
You are categorizing customer support emails as either "billing" or "technical". Examples:
Email: "I was charged twice for last month's invoice."
Category: billing
Email: "The dashboard isn't loading on Chrome since the latest update."
Category: technical
Email: "Can you update my credit card on file?"
Category: billing
Email: [your input here]
Category:
Why it works: the three examples cover both categories with realistic inputs. The format ends with "Category:" so the model's next token will be one of the two category labels.
When it fails: when your categories are not actually mutually exclusive, or when edge cases between the categories are ambiguous. Add a third "neither" or "needs review" category if your real inputs include genuinely ambiguous cases.
Pattern 2 multi-class classification with reasoning
Use when classification needs explanation, when categories are subtle, or when you want auditable output that a human can review.
Prompt structure:
Categorize each piece of feedback into one of: feature request, bug report, praise, complaint, or question. Then explain the categorization in one sentence.
Feedback: "I wish there was a way to bulk-edit projects."
Category: feature request
Reasoning: User is requesting functionality that does not exist.
Feedback: "The export keeps timing out on files larger than 10MB."
Category: bug report
Reasoning: User describes a system failure with specific reproduction details.
Feedback: "Honestly the fastest CRM I've ever used."
Category: praise
Reasoning: Positive evaluation without specific feature request or issue.
Feedback: [your input]
Category:
Why it works: the reasoning step forces the model to justify its classification, which catches misclassifications. The example reasoning is short and specific, modeling the desired output length.
Pattern 3 structured information extraction
Use when pulling structured data out of unstructured text. Common applications include
research synthesis, sales call analysis, and competitive intelligence.
Prompt structure:
Extract the following from each customer interview excerpt: company size, current solution, primary pain point, decision timeline. Return as JSON.
Excerpt: "We're a 200-person SaaS company. We've been using HubSpot for two years but the reporting limitations are killing us. We need to make a decision before our renewal in March."
{ "company_size": "200", "current_solution": "HubSpot", "primary_pain_point": "reporting limitations", "decision_timeline": "March renewal" }
Excerpt: "Just hit 50 employees, scaling fast. Salesforce is too heavy for us right now. We're evaluating options through Q2."
{ "company_size": "50", "current_solution": "Salesforce", "primary_pain_point": "Salesforce too heavy", "decision_timeline": "evaluating through Q2" }
Excerpt: [your input]
Why it works: the JSON structure is consistent across examples. The model will mirror the exact field names and types shown.
When it fails: if your real inputs are missing fields the examples include, the model will sometimes hallucinate a value. Add an example showing how to handle missing fields with null or "not specified."
Pattern 4 style-matched content generation
Use when you need new content that matches an existing voice. Common applications include scaling content production, maintaining brand voice across writers, and persona-specific content adaptation.
Prompt structure:
Below are three examples of LinkedIn posts in our brand voice. Write a new post about [topic] in the same voice.
Example 1: [paste actual high-performing post]
Example 2: [paste actual high-performing post]
Example 3: [paste actual high-performing post]
New post topic: [your topic]
Why it works: the model treats the examples as the operational definition of "our brand voice" rather than relying on stylistic instructions. Three real examples teach voice better than any amount of voice description.
When it fails: if your three examples are stylistically inconsistent, the model produces output that splits the difference. Pick three examples that genuinely represent the voice you want, not three that happen to be recent.
Pattern 5 format transformation
Use when transforming content from one format to another. Common applications include repurposing long-form content into short-form, converting outlines into drafts, and adapting content across channels.
Prompt structure:
Transform each blog article excerpt into a three-tweet thread. Maintain the key claims but adapt for Twitter style and length.
Article: [paste sample article]
Thread: [paste sample three-tweet thread]
Article: [paste second sample]
Thread: [paste second sample thread]
Article: [your input]
Thread:
Why it works: the transformation pattern is shown rather than described. The model learns both the content compression and the platform-specific style adjustments simultaneously.
Pattern 6 sentiment-tagged generation
Use when you need to generate content with specific emotional or rhetorical framing. Common applications include creating contrasting versions of messaging, generating empathy-led vs urgency-led copy, and crafting content for different funnel stages.
Prompt structure:
Write a one-sentence email subject line in each of three styles for a webinar registration: empathetic, curiosity-driven, urgency-driven.
Topic: building your first content calendar
Empathetic: We know content planning feels overwhelming. This webinar will help.
Curiosity-driven: The content calendar mistake 73% of B2B teams make in year one.
Urgency-driven: Last chance to register: Tomorrow's session on content calendars.
Topic: improving lead qualification
Empathetic: [model fills in]
Curiosity-driven: [model fills in]
Urgency-driven: [model fills in]
Why it works: the example demonstrates three distinct rhetorical postures applied to the same topic. The model generalizes the rhetorical patterns to the new topic while preserving the distinctions.
Pattern 7 evaluation and scoring
Use when you need consistent quality scoring across many inputs. Common applications include content audit scoring, lead qualification scoring, and competitive content benchmarking.
Prompt structure:
Score each blog post on the following criteria: keyword optimization (1-5), depth of coverage (1-5), AEO structure (1-5), and call-to-action strength (1-5). Provide a brief justification for each score.
Article excerpt: [sample 1]
Scores: { "keyword_optimization": 4, "depth_of_coverage": 5, "aeo_structure": 2, "cta_strength": 3 }
Justification: { "keyword_optimization": "Primary keyword in H1 and meta", "depth_of_coverage": "Substantive treatment of all subtopics", "aeo_structure": "Missing TL;DR and FAQ", "cta_strength": "CTA present but generic" }
Article excerpt: [sample 2]
Scores: [...] Justification: [...]
Article excerpt: [your input]
Why it works: the example output structure demonstrates that scoring should be both quantitative and justified. The model produces consistent scoring criteria across inputs.
Pattern 8 question generation from content
Use when generating FAQ sections, interview questions, or quiz content. Common applications include AEO-optimized content production and lead-magnet development.
Prompt structure:
Generate ten frequently asked questions a B2B marketing leader might ask after reading the article excerpt. Phrase questions as real buyer queries, not academic phrasings.
Article excerpt: [sample article]
Questions:
1. How much does AEO actually cost to implement?
2. What is the difference between AEO and SEO?
3. [eight more]
Article excerpt: [your input]
Questions:
Why it works: the example questions demonstrate buyer-realistic phrasing rather than the generic "what are the benefits of X" questions models default to.
Pattern 9 outline-to-draft expansion
Use when expanding bullet outlines into full prose. Common applications include scaling content production from approved briefs and converting meeting notes into structured documents.
Prompt structure:
Expand each outline section into a substantive paragraph in the practitioner-direct voice shown in the example.
Outline: "AI Overview optimization fundamentals — Google AI Overviews now appear above blue links — Most informational queries trigger them — Click-through rates drop 30-50%"
Paragraph: "Google AI Overviews changed search the moment they rolled out globally. A buyer searching 'how to build a content calendar' no longer sees ten blue links. They see a 200-word synthesized answer, followed by a list of 3 to 5 source citations, with the classic blue links pushed below the fold. Click-through rates drop 30 to 50 percent on queries where Overviews appear, even for cited sources."
Outline: [your input]
Paragraph:
Why it works: the example shows the level of detail, the voice, and the rhetorical structure (specific example, then claim) in a single demonstration. The model expands new outlines at the same depth and in the same voice.
Pattern 10 prompt chaining setup
Use when a task is too complex for a single prompt and needs multi-step decomposition. Common applications include research synthesis, complex content production, and multi-source analysis.
This pattern is different from the others — it is not a single few-shot prompt but a chain where each step uses few-shot prompting on a specific subtask.
A typical content production chain:
Step 1: few-shot prompt extracts key claims from source material
Step 2: few-shot prompt clusters related claims into themes
Step 3: few-shot prompt drafts a paragraph for each theme
Step 4: few-shot prompt assembles paragraphs into a coherent article in the brand voice
Each step is a separate few-shot prompt with examples specific to that subtask. The output of one step becomes the input to the next. This pattern produces dramatically better results than asking a single prompt to do all four steps at once.
Few-shot Prompting Examples: Common adaptation mistakes
Copying examples without adapting them
These patterns are templates. The wording in any specific example is illustrative. Replace the example content with your own real examples that span your real input space.
Using too few real examples
I wrote two or three examples per pattern in this article for readability. In production, three to five examples almost always outperform two. Add examples until the model's output stabilizes.
Mixing patterns in one prompt
Each pattern works because it is consistent. Mixing classification with extraction with style-matching in a single prompt produces unreliable output. Use prompt chaining (Pattern 10) when you need multiple operations.
Skipping output format examples
If your task needs structured output (JSON, table, specific format), every example output must match that format exactly. The model mirrors what it sees.
Frequently asked questions
How many examples do I need for an effective few-shot prompt?
Three to five examples is the practical range for most tasks. Two examples sometimes work for simple format-matching tasks; six or more rarely add value. The exception is for tasks with high variance in input types — if your real inputs span many different patterns, your examples need to span them too.
Should I copy these few-shot prompting examples directly or adapt them?
Adapt them. The structural shape (delimiters, example count, formatting) is what makes these patterns work. The specific wording in each example should be replaced with content from your actual use case. Copying the structure is the right move; copying the example content is not.
Can I use few-shot prompts in ChatGPT or Claude directly without coding?
Yes. All major chat interfaces (ChatGPT, Claude, Gemini, Perplexity) accept few-shot prompts the same way they accept any other prompt. Paste the full structure including examples and the model handles it. No engineering or API access required for individual use.
What is the best few-shot prompting pattern for content marketing?
For most content marketing teams, the highest-yield patterns are style-matched generation (Pattern 4) for scaling content production in your brand voice, format transformation (Pattern 5) for repurposing content across channels, and question generation (Pattern 8) for producing AEO-optimized FAQ sections.
How do I know if my few-shot prompt is working well?
Test it on inputs where you know the correct answer. Run the prompt on five to ten inputs you have already validated manually. If the model agrees with your judgment, the prompt is reliable. If it disagrees in patterns you can predict, the prompt has structural issues to fix.
Why does my few-shot prompt produce inconsistent results?
Three common causes: examples that are formatted inconsistently (subtle differences in example structure produce inconsistent output structure), examples that do not span the real input space (model fails on input types it never saw), and examples that include the answer to the actual task (rare but happens, model gets confused about what to predict).
Do few-shot prompting patterns work the same in different AI tools?
The structural patterns transfer across all major AI tools — ChatGPT, Claude, Gemini, Llama, Mistral. Specific token counts and minor formatting preferences differ slightly between models. Test the same prompt across the tools you use; expect 80 to 90 percent of the prompt to transfer cleanly with minor adjustments for model-specific quirks.
Should I version control my few-shot prompts?
Yes, especially for production use cases. Few-shot prompts in production are like code — small changes can produce meaningfully different output, and you need to track what changed and why. Most teams use a simple Git repository or a prompt management tool like LangSmith or PromptLayer.
Can few-shot prompts include instructions as well as examples?
Yes, and the strongest prompts usually do both. A clear task instruction at the top, followed by three to five examples, followed by the actual task input. The instruction tells the model what to do; the examples teach it how to do it. Both layers reinforce each other.
How long should each example in a few-shot prompt be?
As long as the realistic task. If your real inputs are 50-word tweets, your examples should be 50-word tweets. If your real inputs are 2,000-word articles, your examples should be 2,000-word articles. The model uses example length as a cue for expected complexity and detail.
What is the difference between few-shot prompting examples and prompt templates?
Few-shot prompting examples are the input-output pairs included inside a prompt. A prompt template is a reusable prompt structure with placeholders for variables (including the few-shot examples themselves). Templates often contain few-shot example slots that get filled in with task-specific examples at runtime.
Can I generate few-shot examples using AI?
Yes, with caution. You can ask a model to generate candidate examples for a task, but always validate them against ground truth before using them in production. AI-generated examples sometimes embed subtle biases or incorrect patterns that propagate through the prompt. Treat AI-generated examples as candidates, not finished work.
Comments