	In Browser
	StumbleUpon
	del.icio.us
	Google
	Google Buzz
	reddit
	LinkedIn

	Facebook
	Twitter
	Linkedin
	E-Mail

Generative AI > Prompt Engineering > Meta-Prompting - Using LLMs to Generate Better Prompts

Meta-Prompting - Using LLMs to Generate Better Prompts

Author: Venkata Sudhakar

Meta-prompting uses an LLM to generate, refine, or evaluate prompts for other LLM tasks. At ShopMax India, instead of manually crafting the perfect prompt for a customer complaint classifier, a meta-prompt asks Claude to write the best possible system prompt for that classifier - then you use the generated prompt in production. This is especially useful when the task domain is well-defined but prompt wording is hard to get right manually.

The process has two steps: a meta-prompt call that describes the task and asks for an optimized prompt, and a validation call that tests the generated prompt on sample inputs. The meta-prompt typically includes the task goal, output format requirements, example inputs and expected outputs, and any constraints. The LLM returns a ready-to-use system prompt that reflects those requirements.

The example below shows ShopMax India using Claude to generate an optimized system prompt for a product review sentiment classifier, then testing the generated prompt on three sample reviews.

import anthropic

client = anthropic.Anthropic()

meta_prompt = """
Write an optimized system prompt for an LLM classifier that will:
- Classify ShopMax India product reviews as: POSITIVE, NEGATIVE, or NEUTRAL
- Extract the main reason for the sentiment in one short phrase
- Return output as JSON: {"sentiment": "...", "reason": "..."}
- Handle Hindi-English mixed reviews (Hinglish)
- Be concise - under 100 words

Provide only the system prompt text, no explanation.
"""

meta_response = client.messages.create(
    model="claude-haiku-4-5",
    max_tokens=256,
    messages=[{"role": "user", "content": meta_prompt}]
)
generated_system_prompt = meta_response.content[0].text
print("Generated system prompt:")
print(generated_system_prompt)
print()

test_reviews = [
    "Ekdum bakwas product hai, Rs 35000 waste ho gaya. Screen broke in 2 days.",
    "Superb TV! Delivered in 2 days to Hyderabad. Picture quality is outstanding.",
    "Packaging was good. Product is ok, nothing special for the price."
]

for review in test_reviews:
    result = client.messages.create(
        model="claude-haiku-4-5",
        max_tokens=64,
        system=generated_system_prompt,
        messages=[{"role": "user", "content": review}]
    )
    print("Review:", review[:50], "...")
    print("Result:", result.content[0].text)
    print()

It gives the following output,

Generated system prompt:
Classify ShopMax India product reviews as POSITIVE, NEGATIVE, or NEUTRAL.
Extract the main sentiment reason in one short phrase.
Respond with JSON only: {"sentiment": "...", "reason": "..."}.
Handle English and Hinglish reviews equally. Be consistent and concise.

Review: Ekdum bakwas product hai, Rs 35000 waste ho gaya. Screen ...
Result: {"sentiment": "NEGATIVE", "reason": "screen broke after 2 days"}

Review: Superb TV! Delivered in 2 days to Hyderabad. Picture quality ...
Result: {"sentiment": "POSITIVE", "reason": "fast delivery and excellent picture quality"}

Review: Packaging was good. Product is ok, nothing special for the pri ...
Result: {"sentiment": "NEUTRAL", "reason": "average product for the price"}

In production at ShopMax India, store the generated prompt in a versioned config file and run it through your full test suite before deploying. Re-run the meta-prompting process whenever the task requirements change - for example, when adding a new sentiment category like MIXED. Combine meta-prompting with automatic evaluation: after generating a prompt, score it against a labeled dataset and only accept it if accuracy exceeds your threshold. This creates a closed-loop prompt optimization pipeline.

Send your comments, suggestions or queries regarding this site to [email protected].