Few-Shot Prompting
The most reliable technique for consistent AI outputs — 3 real templates with copy-paste prompts for Claude, ChatGPT, and any LLM.
Zero-shot vs One-shot vs Few-shot — When to Use Each
| Technique | Examples included | Best for | Risk |
|---|
| Zero-shot | 0 | Simple, well-defined tasks (translate, reformat) | Inconsistent output format |
| One-shot | 1 | Quick tasks where format matters | Model may over-generalize from single example |
| Few-shot ✓ | 2–5 | Classification, extraction, tone matching, scoring | Consumes more tokens; examples must be high quality |
Sentiment Classification (3-shot)
Show the model one example per class before the real input. This locks in the label vocabulary and prevents invented categories.
Classify the sentiment of each customer review as Positive, Negative, or Neutral.
Example 1:
Review: "This product exceeded my expectations. Fast shipping, great quality."
Sentiment: Positive
Example 2:
Review: "Arrived broken. Packaging was a mess. Complete waste of money."
Sentiment: Negative
Example 3:
Review: "It does what it says. Nothing special, but no complaints either."
Sentiment: Neutral
Now classify:
Review: "The interface is intuitive but the mobile app keeps crashing on iOS 17."
Sentiment:
Model output →
Negative
💡 One example per class anchors the label space. The model can't invent "Mixed" or "Frustrated" — it mirrors your exact vocabulary.
JSON Data Extraction (2-shot)
Two examples teach the exact schema. The model learns field names, types, null handling, and the no-preamble output rule from demonstration rather than instruction alone.
Extract structured data from each job posting. Return only valid JSON.
Example 1:
Posting: "Senior React Developer at Acme Corp. 5+ years required. $120k–$160k. Remote OK. Apply by June 1."
Output: {"title":"Senior React Developer","company":"Acme Corp","min_years":5,"salary_min":120000,"salary_max":160000,"remote":true,"deadline":"June 1"}
Example 2:
Posting: "Part-time Data Analyst. No remote. Chicago office. $45/hr. No deadline mentioned."
Output: {"title":"Part-time Data Analyst","company":null,"min_years":null,"salary_min":null,"salary_max":null,"remote":false,"deadline":null}
Now extract:
Posting: "Staff ML Engineer at DeepCo. 8+ years exp. $200k–$250k + equity. Hybrid (NYC). Rolling applications."
Output:
Model output →
{"title":"Staff ML Engineer","company":"DeepCo","min_years":8,"salary_min":200000,"salary_max":250000,"remote":false,"deadline":null}
💡 Example 2 is critical — it teaches null handling for missing fields. Without it, models hallucinate plausible values instead of returning null.
Tone Matching — Brand Voice (3-shot)
When you can't describe brand voice in words, show it. Three examples of actual brand copy teach tone, vocabulary, sentence length, and emotional register more reliably than any style guide.
Rewrite the following product description in our brand voice. Study the examples below to understand our style.
Example 1:
Original: "High-performance running shoes with cushioning technology."
Brand: "Run farther. Hurt less. Our cushioning absorbs impact so your legs don't have to."
Example 2:
Original: "Bluetooth headphones with 40-hour battery life."
Brand: "Forty hours. One charge. Forget cables. Forget limits. Just music."
Example 3:
Original: "Ergonomic office chair with lumbar support."
Brand: "Your back works hard. Give it something that works back. Lumbar support engineered for 8-hour days."
Now rewrite:
Original: "Smart water bottle with hydration tracking and LED reminders."
Brand:
Model output →
"Drink more. Forget less. LED reminders keep you on track — so your body doesn't remind you with a headache at 3pm."
💡 No style guide needed — 3 examples do more work than 500 words of tone documentation.
Few-Shot Best Practices
- Cover the edge cases in your examples — include at least one null/missing-field example for extraction tasks
- Match the real distribution — if 30% of real inputs are negative sentiment, make 1–2 of your 5 examples negative
- Use interleaved human/assistant turns (Claude API) — cleaner than embedding all examples in the system prompt
- Keep examples concise — bloated examples push useful context out of the window
- Vary surface features — don't use three examples that all start with "The product…"; varied phrasing improves generalization
- Put the best example last — the final example before the real input has the most recency weight
→ Auto-optimize my prompt freeFrequently Asked Questions
- What is few-shot prompting?
- Few-shot prompting is a prompt engineering technique where you include 2–5 input→output examples directly in the prompt before presenting the actual input you want the model to process. By showing the model a pattern, it learns the expected output format, tone, and logic without explicit instruction. Few-shot prompting is one of the most reliable techniques for achieving consistent, structured outputs from LLMs like Claude or ChatGPT.
- How many examples should I include in a few-shot prompt?
- 2–5 examples is the practical sweet spot for most tasks. Fewer than 2 examples often fails to establish a clear pattern; more than 5 starts to consume significant context tokens without proportional accuracy gains. For classification tasks, include at least one example per class. For generation tasks (writing, summarization), 3 diverse examples covering different lengths or styles tend to produce the most consistent outputs.
- What is the difference between zero-shot, one-shot, and few-shot prompting?
- Zero-shot: you give only the task instruction, no examples. Works for well-known tasks but produces inconsistent formats. One-shot: one input→output example. Establishes the pattern but can be brittle. Few-shot: 2–5 examples. The most reliable technique — the model picks up the pattern from multiple demonstrations and generalizes it. Few-shot is preferred whenever output format consistency matters (structured data extraction, classification, scoring).
- Does few-shot prompting work with Claude?
- Yes — Claude responds very well to few-shot examples, especially for tasks requiring structured outputs like JSON extraction, classification, or scoring. Claude also supports interleaved human/assistant turns in its API, which is the most natural way to format few-shot examples: alternate human: [input] / assistant: [output] pairs before the final human: [real input] message. This structure is cleaner than embedding all examples in a single system prompt.
- What tasks benefit most from few-shot prompting?
- Few-shot prompting delivers the biggest gains for: (1) Classification — showing the model labeled examples prevents category drift. (2) Structured data extraction — demonstrating the exact JSON schema reduces hallucinated fields. (3) Tone/style matching — showing 3 examples of brand voice teaches it better than describing the voice in words. (4) Scoring/grading — examples calibrate what a 1 vs a 5 looks like. Tasks with well-defined correct answers benefit less; open-ended generation benefits more from system prompt engineering.
Not sure how many examples to include? Let the tool decide.
Paste your draft prompt → the optimizer restructures it and adds examples where they help. Free, BYO Anthropic API key.
→ Improve My PromptRelated Prompt Engineering Guides