System prompt best practices for consistent output formatting

gpt-4o system-prompt formatting

Emma RodriguezFeb 10, 2025

I've spent months refining system prompts for consistent outputs across GPT-4o. Here's what I've learned:

Do's

Put format instructions at the END of the system prompt

Use XML tags to delineate sections: ...

Provide 2-3 examples of desired output

Be explicit about what NOT to include

Don'ts

Don't rely on "be concise" — specify exact word/sentence counts

Don't mix formatting instructions with behavioral instructions

Don't assume the model remembers formatting from previous turns

Template I use

You are a [role]. Your task is [task].

Rule 1
Rule 2

Always respond in this exact format:
[example]

What patterns have worked well for you?

4.1k views24 replies56 likes

3 Replies

Ryan Mitchell Feb 12

This is incredibly helpful. How did you handle failures? Did you re-batch just the failed requests?

Lucia Fernandez Feb 13

Yes, exactly. After each batch completes, I parse the output file and collect any failed request IDs, then create a new batch with just those. Usually only 1-2% fail so the retry batch is small.

failed_ids = [r['custom_id'] for r in results if r['response']['status_code'] != 200]

Robert Chang Feb 14

At our scale (500K+ docs/day), we've also found that splitting by document type gives better results since you can optimize the prompt per type.