The fix is usually to split it. Here's how to tell when, and how.

Signs a prompt is doing two things

• The output has parts that go to different places. The reply goes to the customer; the urgency rating goes into your tracker.
• You regularly delete part of what comes back. If you only wanted the reply today, the rest is noise.
• It has "if" instructions. "If it's a refund request, also..." is often a second prompt hiding inside the first.
• The format rules pull against each other. "Warm and conversational" and "return a rating from 1 to 3" rarely sit well in one answer.

How to split it

1. List what the prompt produces, one output per line.
2. Give each output its own prompt, starting with only the instructions that shape that output.
3. Copy the shared context into both. Who the customer is, what the product does, the tone you write in. A little repetition is cheaper than a tangled prompt.
4. Chain them if one needs the other. Run the first, paste its result into the second.

What each half keeps

Each half keeps the context both need, its own format rules, and its own limits. It drops every instruction that only mattered to the other half. If a line makes sense in both, it probably belongs in both.

For context, we make PromptAlo, a Chrome extension that adds AI metadata tags to prompts you save. A rough heuristic: if the AI tags on a saved prompt miss the point, the wording probably does too.

Have you split a prompt like this, and where did you draw the line?

https://promptalo.conceptgame.co.uk/?utm_source=hashnode&utm_medium=referral&utm_campaign=crosspost-hn-split-saved-prompt

This post was written with AI assistance and edited by the PromptAlo team.

A prompt that moves well between ChatGPT, Claude and Gemini is mostly a prompt that says out loud what one model was quietly filling in. A short way to split it:

Keep in the shared text (model-neutral):

• The task in one plain sentence. No in-jokes from the chat it was born in.
• The format, spelled out. "Five bullets, under 15 words each" travels. "Keep it tight" gets read three different ways.
• The rules as rules. If something must never appear, say so in its own line rather than hoping the tone implies it.
• No app features by name. A prompt that tells one assistant to use a feature only that app has will stall in the other two.

Keep as a short note per model:

• What you had to add. If one model needed an extra "no preamble" line, write that down next to the prompt, not inside it.
• What it does differently. Longer, shorter, more cautious, more formal. One line is plenty.
• Which one you'd reach for first for this particular job, if you have a preference.

The shared text stays clean and portable. The notes hold the model-specific tweaks, so you're not carrying three diverging copies around.

For context, I build PromptAlo, a Chrome extension for keeping prompts that already worked. A saved prompt goes to ChatGPT, Claude or Gemini in one click, which is only useful if the wording holds up in all three.

Which model do you write for first, and what's the first thing you change when you move a prompt to another one?

https://promptalo.conceptgame.co.uk/?utm_source=hashnode&utm_medium=referral&utm_campaign=crosspost-hn-portable-prompts

This post was written with AI assistance and edited by the PromptAlo team.