Polished a prompt for Claude and it falls apart in ChatGPT? Models "hear" instructions differently — here's how to fix a raw prompt and port a finished one to the model you need.
When to use it
Two cases. First — you have the task in your head but can't be bothered to phrase the prompt: a meta-prompt turns your draft into a strong request (role, context, format, criteria). Second — the prompt already works in one model but you need it in another: you port it Claude ↔ ChatGPT ↔ Gemini allowing for their quirks, and pack a frequently used one into a Claude Skill so you stop copy-pasting. The role: prompt engineer.
Prompt: improve a raw request (copy and paste)
You are a prompt engineer. Don't carry out my request — REWRITE it into a strong prompt.
MY DRAFT: "<PASTE YOUR RAW REQUEST>"
Do this:
1. Name what's ambiguous or left unsaid in the draft (as a list).
2. If you can't proceed without answers — ask me at most 3 clarifying questions.
3. Output the improved prompt with these blocks: Role · Context · Task · Output format · Constraints · Criteria for a "good answer".
4. In one line — what exactly you strengthened and why.
Prompt: adapt it to a different model
A finished prompt rarely transfers one to one: what lands with Claude reads differently to ChatGPT. This prompt repackages it for the target model.
You are a prompt engineer. Adapt a finished prompt to a target model: keep the meaning but repackage it around that model's strengths.
SOURCE PROMPT: "<PASTE>"
TARGET MODEL: <Claude / ChatGPT / Gemini>
Do this:
1. Name what in the source is tuned to the previous model and will trip up on the target one.
2. Rewrite the prompt for the target model following its own rules:
- Claude — separate blocks with XML tags (<instructions>, <context>, <input>), give context and "permission to think".
- ChatGPT — put the output format up front, number the steps, state the schema (JSON/table) explicitly.
- Gemini — phrase it directly and briefly, give clear delimiters and sources/data to ground on.
3. Show only the rewritten prompt, plus one line on what you changed and why.
Why exactly this: these are the vendors' own recommendations — Claude is built around XML tags, OpenAI advises structure and an explicit output format, Google advises precision, delimiters and grounding on data. But it isn't magic: compare the answers on your own task anyway.
Pack it into a Claude Skill (SKILL.md)
If you run these prompts often, turn them into a Skill (Agent Skill) so Claude picks them up itself, no copy-pasting. A skill is a folder with a SKILL.md file: YAML at the top with two required fields (name and description), the instruction below. Claude matches your request against the description specifically, so write both what the skill does and when to apply it. In Claude Code, put the folder in ~/.claude/skills/ or .claude/skills/; on claude.ai, upload a zip in Settings → Features (you need a plan with code execution enabled). The skeleton:
---
name: prompt-porter
description: Improves a raw request and ports a finished prompt to a target model (Claude/ChatGPT/Gemini). Use when asked to "improve my prompt" or "rewrite this prompt for another model".
---
## Instruction
(both prompts above go here)
Filled-in example
Porting: the source prompt is tuned for ChatGPT — "Return JSON matching this schema…", dry numbered steps. Target model — Claude. What the AI typically returns: (1) it points out that a rigid JSON schema without context is overkill for Claude and that "permission to think" isn't granted; (2) it wraps the task and the data in
Pro tips
- The most valuable part of the meta-prompt is the clarifying-questions block: it pulls out what you forgot to specify.
- When porting, don't copy phrasings verbatim: what works as "permission to think" for Claude is better compressed for Gemini.
- A skill is triggered by its description. If Claude "doesn't see" it — make the description more specific and add trigger phrases (the way a user would actually ask).
- For the work itself use the strongest model available (Opus/GPT class): the quality of the rewritten prompt depends on how much the model has seen.