Lab 3 — Prompt Engineering¶

⏱️ ~25 min

Turn a vague prompt into a reliable one: zero-shot → few-shot → structured output.

In [ ]:

import os, getpass
PROVIDER = "openai"  # or "anthropic"
if PROVIDER == "openai":
    !pip install -q openai
    if not os.environ.get("OPENAI_API_KEY"):
        os.environ["OPENAI_API_KEY"] = getpass.getpass("OpenAI API key: ")
else:
    !pip install -q anthropic
    if not os.environ.get("ANTHROPIC_API_KEY"):
        os.environ["ANTHROPIC_API_KEY"] = getpass.getpass("Anthropic API key: ")

def chat(messages, temperature=0.7, model=None):
    if PROVIDER == "openai":
        from openai import OpenAI
        client = OpenAI()
        resp = client.chat.completions.create(
            model=model or "gpt-4o-mini", messages=messages, temperature=temperature)
        return resp.choices[0].message.content
    else:
        from anthropic import Anthropic
        client = Anthropic()
        system = "".join(m["content"] for m in messages if m["role"] == "system")
        turns = [m for m in messages if m["role"] != "system"]
        resp = client.messages.create(
            model=model or "claude-3-5-haiku-20241022", max_tokens=1024,
            temperature=temperature, system=system or None, messages=turns)
        return resp.content[0].text

import os, getpass PROVIDER = "openai" # or "anthropic" if PROVIDER == "openai": !pip install -q openai if not os.environ.get("OPENAI_API_KEY"): os.environ["OPENAI_API_KEY"] = getpass.getpass("OpenAI API key: ") else: !pip install -q anthropic if not os.environ.get("ANTHROPIC_API_KEY"): os.environ["ANTHROPIC_API_KEY"] = getpass.getpass("Anthropic API key: ") def chat(messages, temperature=0.7, model=None): if PROVIDER == "openai": from openai import OpenAI client = OpenAI() resp = client.chat.completions.create( model=model or "gpt-4o-mini", messages=messages, temperature=temperature) return resp.choices[0].message.content else: from anthropic import Anthropic client = Anthropic() system = "".join(m["content"] for m in messages if m["role"] == "system") turns = [m for m in messages if m["role"] != "system"] resp = client.messages.create( model=model or "claude-3-5-haiku-20241022", max_tokens=1024, temperature=temperature, system=system or None, messages=turns) return resp.content[0].text

1. The vague prompt (bad)¶

In [ ]:

review = "the app keeps crashing when I upload a big file and im on wifi"
print(chat([{"role": "user", "content": f"Classify this: {review}"}]))

review = "the app keeps crashing when I upload a big file and im on wifi" print(chat([{"role": "user", "content": f"Classify this: {review}"}]))

2. Add a system prompt + clear task (better)¶

In [ ]:

system = {"role": "system", "content":
    "You are a support-ticket triager. Classify the ticket's category and severity."}
print(chat([system, {"role": "user", "content": review}]))

system = {"role": "system", "content": "You are a support-ticket triager. Classify the ticket's category and severity."} print(chat([system, {"role": "user", "content": review}]))

3. Few-shot — show the exact shape you want¶

In [ ]:

few_shot = [
    {"role": "system", "content": "Classify support tickets. Reply exactly like the examples."},
    {"role": "user", "content": "login button does nothing on safari"},
    {"role": "assistant", "content": "category=auth; severity=high"},
    {"role": "user", "content": "the font on the about page looks a bit small"},
    {"role": "assistant", "content": "category=ui; severity=low"},
    {"role": "user", "content": review},
]
print(chat(few_shot, temperature=0))

few_shot = [ {"role": "system", "content": "Classify support tickets. Reply exactly like the examples."}, {"role": "user", "content": "login button does nothing on safari"}, {"role": "assistant", "content": "category=auth; severity=high"}, {"role": "user", "content": "the font on the about page looks a bit small"}, {"role": "assistant", "content": "category=ui; severity=low"}, {"role": "user", "content": review}, ] print(chat(few_shot, temperature=0))

4. Structured output — JSON your code can parse¶

In [ ]:

import json
json_prompt = [
    {"role": "system", "content":
     'Return ONLY valid JSON: {"category": str, "severity": "low|medium|high"}. No prose.'},
    {"role": "user", "content": review},
]
raw = chat(json_prompt, temperature=0)
print(raw)
print("Parsed:", json.loads(raw))

import json json_prompt = [ {"role": "system", "content": 'Return ONLY valid JSON: {"category": str, "severity": "low|medium|high"}. No prose.'}, {"role": "user", "content": review}, ] raw = chat(json_prompt, temperature=0) print(raw) print("Parsed:", json.loads(raw))

Takeaway: each step removed ambiguity. Few-shot + structured output makes the model's response safe to feed into code.

✅ Did it work?¶

• The few-shot step produced the exact category=...; severity=... shape from the examples.
• The final JSON step printed valid JSON and json.loads parsed it into a dict with no error.