Advanced Prompting Techniques

Beyond basic instructions: reasoning elicitation, few-shot calibration, system prompt architecture, and pipeline decomposition.

Learning Objectives

• Apply chain-of-thought prompting to improve accuracy on complex tasks
• Construct effective few-shot examples that guide model behaviour
• Design robust system prompts that define scope and persona
• Use XML tags to structure inputs and outputs for Claude
• Break complex tasks into prompt chains for better reliability

Chain-of-Thought Prompting

Chain-of-thought (CoT) prompting asks the model to reason step by step before giving its final answer. Research consistently shows this improves accuracy on arithmetic, logic, and multi-step problems.

The simplest trigger phrase: "Let's think step by step." — but explicit step structures work even better.

prompt = """Solve the following problem step by step, showing your reasoning.
After reasoning, state your final answer on a new line starting with "Answer:".

Problem: A store has 240 apples. They sell 3/8 of them in the morning and
1/4 of the remainder in the afternoon. How many apples are left?

Let's think step by step:"""

Few-Shot Prompting

Providing example input → output pairs inside the prompt calibrates the model's response format, tone, and reasoning style without any fine-tuning.

Example Structure

Classify the sentiment of customer reviews as POSITIVE, NEUTRAL, or NEGATIVE.

Review: "Delivery was lightning fast and the product exceeded expectations!"
Sentiment: POSITIVE

Review: "It works fine. Nothing special."
Sentiment: NEUTRAL

Review: "Arrived broken and customer service was unhelpful."
Sentiment: NEGATIVE

Review: "{{USER_REVIEW}}"
Sentiment:

Selection Criteria

• Diverse: cover edge cases, not just the easy majority class
• Representative: similar distribution to real inputs
• Count: 3–8 examples is usually optimal; more rarely helps and costs tokens
• Consistent format: inconsistent examples confuse the model more than no examples

System Prompts

The system prompt is your persistent configuration layer. It scopes every turn of the conversation and is prepended before the user's first message.

What to Put in System Prompts

• Persona: "You are Aria, a friendly customer service agent for AcmeCorp."
• Hard constraints: "Never discuss competitor products. Decline politely if asked."
• Output conventions: "Always respond in the same language the user writes in."
• Knowledge scope: "You have access to our product catalogue below. Do not speculate beyond it."

Example System Prompt

You are a senior code reviewer specialising in Python and security.
Your job is to review code snippets for bugs, security vulnerabilities, and style issues.

Rules:
- Always list issues by severity: Critical → High → Medium → Low
- Suggest a concrete fix for every issue you raise
- Do not comment on correct code; only flag problems
- If the code is safe and idiomatic, respond only: "LGTM — no issues found."

XML Tags in Claude

Claude is trained to pay special attention to XML-like tags. Using them creates unambiguous structure that separates instructions, data, and expected outputs.

Why XML Works Well

• Prevents prompt injection — user data can't escape its <document> wrapper
• Creates clear delineation of instructions vs content
• Claude can be asked to wrap its own output in tags for easy parsing

prompt = """<instructions>
Extract all action items from the meeting notes below.
Return them as a JSON array of objects with fields: owner, action, due_date.
If a field is not mentioned, use null.
</instructions>

<document>
{meeting_notes}
</document>

<output>
"""
# Claude will fill in and close the output tag

Prompt Chaining

Complex tasks rarely fit in one prompt. Prompt chaining breaks them into sequential steps where each output feeds the next — improving reliability and making each step independently testable.

def chain_summarise_then_classify(document: str) -> dict:
    # Step 1: summarise
    summary = client.messages.create(
        model="claude-haiku-4-5", max_tokens=300,
        messages=[{"role": "user", "content":
            f"Summarise in 3 sentences:\n\n{document}"}]
    ).content[0].text

    # Step 2: classify the summary
    label = client.messages.create(
        model="claude-haiku-4-5", max_tokens=20,
        messages=[{"role": "user", "content":
            f"Classify as SPORTS/POLITICS/TECH/OTHER:\n\n{summary}"}]
    ).content[0].text.strip()

    return {"summary": summary, "label": label}

Demo

Live examples showing CoT reasoning and XML extraction.

Lab & Homework