প্রম্পট ইঞ্জিনিয়ারিং

একই LLM, একই question — দুজন user এর কাছে দুই রকম answer। কেন? Prompt লেখার তফাৎ। 'Write a poem' vs 'Write a 4-line haiku about autumn rain in Tagore style' — দ্বিতীয় টায় quality আকাশ-পাতাল ভালো। Prompt Engineering — AI কে effectively instruct করার বিজ্ঞান এবং art।

Prompt Engineering = LLM থেকে best output বের করার জন্য input design করার practice। Core techniques: Zero-shot (শুধু instruction), Few-shot (কিছু example), Chain-of-Thought (step by step think কর), Role assignment ('You are an expert...'), Constraint specification (format, length, style), Self-consistency (multiple sample, majority pick), ReAct (reason + act loop)।

ভাবুন employee কে instruction দেওয়া — vague বললে ('do something') যা ইচ্ছা করবে। Specific বললে ('write a 300-word product description focused on durability, target audience 30-50 age, in friendly tone, end with CTA') — exact যা চান পাবেন। LLM ও তাই — যত clear, structured, contextual instruction, তত predictable + high-quality output।

• ChatGPT custom GPT — system prompt এ অনেক engineering।
• Cursor/Copilot — code completion prompt internally optimized।
• Marketing copy generation — brand voice prompt template।
• Customer support bot — persona + escalation rule প্রম্পট এ।
• Coding agent — ReAct prompting + tool use।

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://ai.gateway.lovable.dev/v1",
    api_key=os.environ["LOVABLE_API_KEY"],
)

def ask(messages, label):
    res = client.chat.completions.create(
        model="google/gemini-2.5-flash",
        messages=messages,
        temperature=0.2,
    )
    print(f"\n--- {label} ---")
    print(res.choices[0].message.content)

problem = "If a shop sold 240 apples on Monday and twice as many on Tuesday, how many in total?"

ask([{"role": "user", "content": problem}], "Zero-shot")

ask([
    {"role": "system", "content": "You are a math tutor. Think step by step before giving the final answer."},
    {"role": "user", "content": problem},
], "Chain-of-Thought")

few_shot = [
    {"role": "system", "content": "Convert sentence to JSON with fields: subject, action, object."},
    {"role": "user", "content": "Sara reads the book."},
    {"role": "assistant", "content": '{"subject": "Sara", "action": "reads", "object": "the book"}'},
    {"role": "user", "content": "The cat chased the mouse."},
]
ask(few_shot, "Few-shot Structured Output")

• Vague instruction — 'write something good' মানে কী?
• Constraint না দেওয়া (length, format) — output unpredictable।
• Few-shot এ inconsistent example — model confused।
• System আর user message gulier।

1. একটা task এ zero-shot vs few-shot accuracy compare।
2. JSON output prompt লিখুন, parse করে validate।
3. Chain-of-Thought দিয়ে math problem solve।
4. Role prompt ('You are a senior lawyer...') effect দেখুন।

• Prompt Engineering = LLM থেকে best output বের করার design discipline।
• Zero-shot, Few-shot, CoT, Role, Format — core techniques।
• Clear, specific, structured prompt = predictable output।
• Iteration ছাড়া perfect prompt হয় না।
• Modern AI engineer এর core skill।