Day 9: Advanced Prompt Engineering — Few-Shot, Chain-of-Thought and Structured JSON

Prompt engineering is not just talking to a chatbot. It is the programmable interface that lets you get dependable, repeatable results from an AI, and in an automated system, dependable is the whole point.

In plain terms: Imagine briefing a new colleague. "Sort these tickets" gets you a guess. A short brief, a couple of worked examples, and a note on how you want the result written down gets you what you actually need. A good prompt is that brief.

1. Few-shot prompting

With zero-shot prompting you give only an instruction. With few-shot prompting you add two or three perfectly formatted examples of the input and the output you want. The examples ground the model: it copies the pattern, so the answers come back in the same shape every time.

2. Chain-of-thought (CoT)

Chain-of-thought prompting asks the model to write out its reasoning step by step before it gives the final answer. If you ask an AI to analyse the dividend yield of a real-estate trust, making it do the maths in order, instead of jumping to a number, greatly reduces logic errors. A simple way to ask is: "Work through this step by step, then give the final answer on a new line."

3. Structured JSON output

In automated workflows, free text is not much use. If the answer must go into a database or another program, you have to constrain the prompt so the model returns a strict JSON structure that code can parse. Many AI platforms now also offer a built-in structured-output mode that enforces a schema.

Here is a small, runnable example that builds a few-shot prompt and then checks the reply before trusting it. (It does not call a model: it shows what you send and how you validate what comes back.)

import json

# 1. A few-shot prompt: two worked examples, then the new ticket
examples = [
    ("Disk is full on the file server", {"team": "hardware", "priority": "medium"}),
    ("Cannot log in after the password reset", {"team": "identity", "priority": "high"}),
]
prompt = "Classify each IT ticket. Reply with JSON only, using keys team and priority.\n\n"
for text, answer in examples:
    prompt += f'Ticket: "{text}"\nAnswer: {json.dumps(answer)}\n\n'
prompt += 'Ticket: "VPN keeps dropping every few minutes"\nAnswer:'
print(prompt)

# 2. Whatever the model replies, check it before your code trusts it
def parse_ticket(reply: str) -> dict:
    data = json.loads(reply)                          # fails if the reply is not valid JSON
    assert set(data) == {"team", "priority"}, "unexpected keys"
    assert data["priority"] in {"low", "medium", "high"}, "bad priority"
    return data

print(parse_ticket('{"team": "network", "priority": "high"}'))     # accepted
try:
    parse_ticket("Sure! The ticket looks like a network issue.")   # chatty text is rejected
except Exception as err:
    print("rejected:", type(err).__name__)

Even with a good prompt, a model can occasionally break the format, so the checking step matters.

Choosing the technique

Coming Up Next

Day 10: Tokenization, context windows and cost optimization.

#PromptEngineering #LLMs #AIImplementation #SoftwareEngineering #Automation