Few Shot And Zero Shot Prompting

Why this matters

As a Prompt Engineer, you often need reliable outputs without retraining a model. Zero-shot and few-shot prompting let you steer an LLM to classify, extract, transform, and reason—using only instructions and a handful of examples. Mastering these patterns reduces iteration time, improves consistency, and saves tokens.

• Customer support: route tickets by intent and urgency.
• Data pipelines: extract structured fields (names, dates, amounts) into JSON.
• Content operations: rewrite text to a brand voice or format, on demand.
• Analytics: summarize findings and enforce concise output shapes.

Concept explained simply

Zero-shot: You describe the task and constraints without providing examples. The model infers the pattern from your instruction.

Few-shot: You give a small set of example pairs (input → output) that demonstrate the exact style, format, and level of detail you want. The model imitates that pattern for new inputs.

Mental model

Think of the model as a pattern-completer. Your prompt defines the pattern. Examples are gravity: they pull the output toward demonstrated labels, formats, and tone. The most recent examples typically influence the most. Keep labels explicit, formats rigid, and instructions unambiguous.

• Frame: Task → Rules → Format → (Examples) → New input.
• Delimit inputs with clear markers (e.g., triple dashes or XML-like tags).
• Name your labels and fields exactly once; repeat them consistently.

Reusable prompt templates

Task: Classify the text into one of: {LABELS}. Output only the label.
Rules:
- Use exactly one label from {LABELS}.
- No extra words.

Text:
---
{TEXT}
---

Answer:

Task: Extract data into strict JSON. If a field is missing, use null.
Format (single line JSON): {"field_a":"string","field_b":"string","amount":"number or null"}

Examples:
Input: "{E1_INPUT}"
Output: {E1_JSON}

Input: "{E2_INPUT}"
Output: {E2_JSON}

Now extract from the new input:
Input: "{NEW_INPUT}"
Output:

Task: Solve the problem and give a concise final answer.
Rules:
- First, list up to 2 short reasoning steps.
- Then output: Final answer: <value>

Problem:
---
{PROBLEM}
---

Worked examples

Example 1: Zero-shot labeler (support emails)

Task: Label the email as one of: billing, technical, cancellation, other. Output only the label.
Email:
---
I was charged twice this month. Please fix.
---
Answer:

Example 2: Few-shot JSON extractor (event details)

Task: Extract event details into JSON: {"title":"string","date":"YYYY-MM-DD","time":"HH:MM","city":"string or null"}

Examples:
Input: "Jazz Night on Feb 14 at 7pm, downtown Chicago"
Output: {"title":"Jazz Night","date":"2024-02-14","time":"19:00","city":"Chicago"}

Input: "Workshop: Prompting 101, 2024-03-02 09:30 (online)"
Output: {"title":"Prompting 101","date":"2024-03-02","time":"09:30","city":null}

Now extract:
Input: "Meetup – Data Friday, March 8 at 6:15pm, Seattle"
Output:

Example 3: Few-shot style transfer (brand tone)

Task: Rewrite in the brand voice: friendly, concise, action-oriented. Max 2 sentences.

Examples:
Original: "Our application has undergone maintenance."
Rewrite: "We just finished maintenance—thanks for your patience!"

Original: "Your request will be processed within 3 business days."
Rewrite: "We’re on it—expect an update within 3 business days."

Now rewrite:
Original: "The report has been generated successfully."
Rewrite:

Choosing zero vs. few-shot

Token and evaluation tips

• Keep examples short. Trim fluff and keep just the essential pattern.
• Order matters: put the most relevant, recent example closest to the new input.
• Lock output shape: state the exact schema and “Output only the JSON. No extra text.”
• Test variation: try 3–5 diverse inputs to catch edge cases.
• Measure: check accuracy, format validity, and consistency across runs.

Exercises

Do these now. They mirror the graded exercises below.

Exercise 1: Zero-shot labeler for app reviews

Goal: Classify each review as one of: bug, feature_request, praise, other. Output only the label.

Reviews:
1) "Crashes every time I open settings."
2) "Love the new dark mode!"
3) "Could you add export to CSV?"

Task: Classify the text into one of: bug, feature_request, praise, other. Output only the label.

Text:
---
Crashes every time I open settings.
---
Answer: bug

Text:
---
Love the new dark mode!
---
Answer: praise

Text:
---
Could you add export to CSV?
---
Answer: feature_request

Exercise 2: Few-shot JSON extractor for job posts

Goal: Extract {"title":"string","company":"string","location":"string or null","salary_range":"string or null"}. If missing, use null. Output only one JSON object per input.

Examples:
Input: "Acme seeks Data Analyst in Berlin. €55–70k."
Output: {"title":"Data Analyst","company":"Acme","location":"Berlin","salary_range":"€55–70k"}

Input: "Remote Backend Engineer at Nimbus (US only)."
Output: {"title":"Backend Engineer","company":"Nimbus","location":"Remote (US only)","salary_range":null}

New input:
"Orion hiring Product Manager, London, competitive pay"

Task: Extract fields into JSON exactly as: {"title":"string","company":"string","location":"string or null","salary_range":"string or null"}. Output only JSON.

Examples:
Input: "Acme seeks Data Analyst in Berlin. €55–70k."
Output: {"title":"Data Analyst","company":"Acme","location":"Berlin","salary_range":"€55–70k"}

Input: "Remote Backend Engineer at Nimbus (US only)."
Output: {"title":"Backend Engineer","company":"Nimbus","location":"Remote (US only)","salary_range":null}

Now extract:
Input: "Orion hiring Product Manager, London, competitive pay"
Output: {"title":"Product Manager","company":"Orion","location":"London","salary_range":null}

Checklist before you move on

• Your prompts state the exact label set or JSON schema.
• Outputs are single-label or strict JSON with no extra text.
• For missing data, you used null consistently.

Common mistakes (and fixes)

• Ambiguous label names. Fix: define labels clearly, mutually exclusive.
• Free-form outputs when a schema is needed. Fix: show the schema and say “Output only JSON”.
• Too many or too long examples. Fix: 2–4 concise shots are usually enough.
• Examples that contradict rules. Fix: ensure every example follows the stated format.
• Leaky instructions inside examples. Fix: keep examples generic; avoid sensitive data.

How to self-check your prompts

• Does the model ever produce extra prose? If yes, tighten the “Output only …” rule.
• Do labels drift? Add a counter-example or clarify boundaries.
• Are rare cases handled? Add one targeted example for the edge case.
• Is the format always valid JSON? Validate with a JSON parser and shorten examples if errors rise.

Practical projects

1. Support ticket router: zero-shot classify tickets into product_areas and priority. Add few-shot examples for tricky categories.
2. Invoice line-item extractor: few-shot to extract vendor, date, total, and currency into JSON. Include one example with missing currency.
3. Knowledge base summarizer: few-shot to summarize articles into 3 bullets and a one-line takeaway, enforcing a strict bullet format.

Who this is for

• Aspiring prompt engineers and analysts working with LLMs.
• Developers adding LLM features to apps.
• Content and ops teams needing repeatable transformations.

Prerequisites

• Basic familiarity with LLM prompts and token limits.
• Comfort reading and writing simple JSON.

Learning path

1. Start with zero-shot classification and simple extraction.
2. Add 2–3 examples to fix inconsistencies (few-shot).
3. Enforce strict output formats (schemas, label sets).
4. Test on diverse inputs and refine with minimal examples.

Next steps

• Introduce anti-examples for common pitfalls (what not to do).
• Experiment with ordering of examples to see impact on outputs.
• Build a small library of reusable templates for your domain.

Mini challenge

Create a 2-shot prompt that converts informal meeting notes into a JSON agenda: {"title":"","date":"YYYY-MM-DD","actions":["..."]}. Include one example with no actions (empty array). Test on a new note.

Task: Convert notes to JSON agenda: {"title":"string","date":"YYYY-MM-DD","actions":["string"]}. If no actions, use an empty array. Output only JSON.

Examples:
Notes: "Team sync 2024-05-02: review Q1, assign Alice to draft report"
Output: {"title":"Team sync","date":"2024-05-02","actions":["Review Q1","Assign Alice to draft report"]}

Notes: "Budget chat 2024-06-10: discussed approach, no assignments"
Output: {"title":"Budget chat","date":"2024-06-10","actions":[]}

Now convert:
Notes: "Project Orion 2024-07-03: finalize scope; Ben to contact vendor"

Quick test

When you’re ready, take the quick test below. Available to everyone; only logged-in users get saved progress.