Zero-shot Prompting

Zero-shot prompting means you give the model no examples -- just clear instructions and let it do its thing. This section's goal: help you write reusable, stable, and verifiable zero-shot prompts.

---

Learning Path (Suggested Order)

1. Beginner: Nail the minimal structure of "instruction + output format"
2. Intermediate: Add constraints and fallbacks for edge cases
3. Practical: Break business tasks into verifiable conditions

---

What Is Zero-Shot Prompting?

Zero-shot prompting is simple: no examples, just rules. It works best when the task is clear, labels are well-defined, and the output format is straightforward.

┌─────────────────────────────────────────────────────────────┐
│                    Zero-shot Prompt Flow                      │
├─────────────────────────────────────────────────────────────┤
│                                                             │
│   Task Description →  Labels/Constraints → Input Data → Output │
│   (what to do)        (how to do it)      (text/data)   (structured) │
│                                                             │
└─────────────────────────────────────────────────────────────┘

---

Why Does Zero-Shot Matter?

---

Business Output (PM Perspective)

You can ship these quickly with zero-shot prompts:

• Minimum Viable Feature (MVP)
• Structured Output Templates (ready for automation)
• Quantifiable Validation (accuracy / consistency)

Completion criteria (suggested):

• Read this page + finish 1 exercise + do 1 self-check

---

Core Prompt Structure

The key to zero-shot prompts: task description + output constraints.

Goal: What task to perform
Labels/Rules: Options and boundaries
Format: Output structure
Input: Content to process

General Template

Complete the following task:
{task}

Rules:
{rules}

Output format:
{output_format}

Input:
{input}

---

Quick Start: Sentiment Classification

Prompt:

Classify the text as neutral, negative, or positive.

Text: I think this vacation was okay.
Sentiment:

Output:

neutral

Note: No examples were provided. That's the whole point of zero-shot.

---

Example 1: Intent Recognition

Scenario: Customer service messages

Identify the user's intent. Only output one of: inquiry / complaint / refund / other.

User message: The headphones I bought broke after two days. How are you going to handle this?

Intent:

---

Example 2: Information Extraction

Scenario: Extract key info from text

Extract name, company, and title from the text.
Output JSON: {"name":"","company":"","title":""}

Text: I'm Alice, and I work as a Product Manager at JR Academy.

---

Example 3: Formatted Summary

Scenario: Meeting notes

Organize the following content into 3 bullet points, each no longer than 20 words.

Content: We decided to finish requirement breakdown this week and start development next week. A handles the backend API, B handles the frontend pages.

---

Migration Template (Swap Variables to Reuse)

Task: {task}
Rules: {label_space_or_rules}
Output: {format}
Input: {input_text}

---

Self-Check Checklist (Before Submitting)

• Is the task clear enough?
• Is the output format fixed?
• Is there a fallback (output unknown when unsure)?
• Can you reproduce stable results within 3 tries?

---

Advanced Tips

1. Restrict output: State "only output the label/JSON".
2. Add fallbacks: If unsure -> unknown.
3. Define boundaries: Specify positive/negative example keywords or rules.
4. Control temperature: temperature=0 for better consistency.
5. Step-by-step prompting: Break complex tasks into two zero-shot steps.

---

Common Problems & Solutions

---

Latest Research at a Glance

• Instruction Tuning: Fine-tuning on multi-task instruction data significantly improves zero-shot generalization, especially on unseen tasks.
• RLHF/Instruct: Aligning model behavior through human feedback improves the reliability and consistency of "follow the instruction" outputs.
• Prompt Optimization: Treating the prompt itself as an optimizable object -- iteratively generating better instructions to boost zero-shot performance.

---

API Call Examples

Python (OpenAI)

from openai import OpenAI

client = OpenAI()

def zero_shot_classify(text: str) -> str:
    response = client.chat.completions.create(
        model="gpt-4o",
        messages=[
            {
                "role": "system",
                "content": "You are a classifier. Only output labels."
            },
            {
                "role": "user",
                "content": f"Classify: positive/negative/neutral\nText: {text}\nLabel:"
            }
        ],
        temperature=0,
        max_tokens=10
    )
    return response.choices[0].message.content.strip()

Python (Claude)

import anthropic

client = anthropic.Anthropic()

def zero_shot_classify(text: str) -> str:
    message = client.messages.create(
        model="claude-sonnet-4-20250514",
        max_tokens=10,
        messages=[
            {
                "role": "user",
                "content": f"Classify: positive/negative/neutral\nText: {text}\nLabel:"
            }
        ]
    )
    return message.content[0].text.strip()

---

Hands-On Exercises

Exercise 1: Classification

Classify the following reviews as positive / negative / neutral:
1. Shipping was super fast, great experience
2. Quality is so-so
3. Completely unusable, so disappointed

Exercise 2: Extraction

Extract the date and amount from the text, output as JSON.
Text: The order will ship on 2025-03-12, total price is $129.

Exercise 3: Summarization

Summarize the following content into 2 action items.
Text: This week we need to finish the page redesign and API integration. Testing is scheduled for Friday.

---

Exercise Scoring Rubric (Self-Assessment)

---

Related Reading

• Few-shot Prompting
• Classification Prompts
• Prompt Elements
• Model settings

---

References

• https://arxiv.org/pdf/2109.01652.pdf
• https://arxiv.org/abs/1706.03741
• https://arxiv.org/abs/2203.02155
• https://arxiv.org/abs/2309.03409

---

Summary

1. Zero-shot prompting works best for clear tasks and quick validation.
2. Fixing the output format dramatically improves stability.
3. You must define boundaries and fallbacks.
4. Low temperature = more stable, more consistent.
5. Templates make your prompt workflow reusable.