GPT-4.1 not respecting single line break between bold title and paragraph

Hi everyone,

I’m trying to enforce a very specific text formatting rule when working with GPT-4.1, but the model doesn’t consistently follow it—despite very clear and repeated instructions.

Here’s the structure I want GPT-4.1 to follow:

— Each section starts with a bold title on its own line.
— The paragraph should begin immediately after, with exactly one line break (1x Enter) between the title and the paragraph—no extra blank line, and the title and paragraph must not appear on the same line.

I’m using this format to present content in clean, readable blocks (like mini text “cards” with title + paragraph). It’s especially important for writing with clear structure—both for human readability and for post-processing.

I’ve tried:

— Explaining the rule with detailed prompts.
— Providing correct and incorrect examples.
— Rephrasing the request in multiple formats (Markdown, plain text, lists, etc.).

Still, GPT-4.1:

— Adds an extra blank line between the title and paragraph, or
— Places the title and paragraph on the same line.

My questions:

:backhand_index_pointing_right: Is there a prompt structure or system message that reliably enforces this format?

:backhand_index_pointing_right: Or is this a known limitation of GPT-4.1’s text generation or rendering behavior?

Thanks in advance for any tips or workarounds.
Gus

2 Likes

Looks like the model is too overfitted on post training to obey any system instructions. Just be thankful it is not like o3, producing dumb tables and em-spaces as code indentation and filling code with multi-byte code points.

You have a short window to enjoy gpt-4.5 to follow instructions before they flip the off switch on the only AI that is powered by IQ.

You could probably write up a regex for this. Even just replacing \n\n with \n if you want it to hit the entire document.

I do not relent, another style of system message. Then a multi-shot from GPT-4.5.

Gets us close to your need:

Playground preset, you can copy out code and adapt the messages. Then see how long it persists.

https://platform.openai.com/playground/p/15YxoD7xSnrhkrXEqj3oP93X?mode=chat