What Is Meta-Prompting?
Prompts that produce other prompts: the recursive power of LLMs.
Meta-Prompting Defined
Meta-prompting is the practice of writing prompts whose output is other prompts. Instead of directly solving a task, a meta-prompt instructs the model to generate the instructions that will then solve the task. It is a layer of abstraction above regular prompting.
First-Order vs. Meta-Prompting
The distinction between first-order and meta-prompting: first-order prompts produce task outputs (summaries, code, analysis). Meta-prompts produce prompts, evaluation criteria, or system instructions that can then be applied to the actual task.
import anthropic
client = anthropic.Anthropic(api_key='YOUR_API_KEY')
# FIRST-ORDER prompt (produces a task output directly)
first_order = 'Write a customer service response for a user whose order was delayed.'
# META-PROMPT (produces a prompt that can then solve similar tasks)
meta_prompt = (
'Design a system prompt for a customer service AI agent '
'that handles order delay complaints. The agent should '
'be empathetic, solution-focused, and proactively offer '
'compensation when appropriate. Output only the system prompt.'
)
response = client.messages.create(
model='claude-opus-4-5', max_tokens=800,
messages=[{'role': 'user', 'content': meta_prompt}]
)
generated_system_prompt = response.content[0].text
print('Generated system prompt:')
print(generated_system_prompt[:300], '...')