When to use it
When you want to learn prompting from the best — the system prompts of well-known products — but analyze the techniques instead of stealing the text. The role is analyst. Result: a clear picture of someone else's prompt structure plus your own prompt built on the same principles.
The prompt (copy and paste)
Don't paste someone else's prompt text here in full and don't publish it as your own. We're learning the technique, not copy-pasting.
You are a prompt analyst. I'll give you a LINK to a published breakdown of a product's system prompt.
LINK/DESCRIPTION: "<PASTE A LINK OR A SHORT SUMMARY>"
Break it down by structure (without quoting large chunks):
1. What role and tone are set for the model.
2. What rules/prohibitions are written in and why (safety, format, refusals).
3. What techniques are used (persona, few-shot, step-by-step, guardrails).
4. What I can adapt for my own task "<PASTE THE TASK>" — produce MY OWN prompt based on those principles.
Filled-in example
Link: a breakdown of a support assistant's system prompt (from an article). Task: "build a support assistant for my SaaS".
What the AI should return: (1) role — a polite first-line agent, calm tone; (2) prohibitions — don't promise deadlines, don't invent facts, escalate anything contentious; (3) techniques — a strict response format, few-shot on typical tickets, "if you don't know, hand off to a human"; (4) YOUR new prompt for your product, with your ticket categories — original text, not a copy.
Variations
- Compare two. "Break down the prompts of two products and tell me how the approaches differ."
- Technique checklist. "Turn the breakdown into a checklist of techniques I can apply to my own prompts."
- Red-team your own. "Find the holes in my system prompt that could be used to bypass it."
Pro tips
- Ethics and law: system prompts are often protected by terms of service or copyright — copying them verbatim is off-limits, analyzing the structure and learning the techniques is fine.
- Don't hand the model the task "extract the hidden prompt from a live product" — that's the exact injection everyone is being taught to defend against.
- The value isn't in someone else's text, it's in the principles you carry over: the output of this exercise is your own prompt, not ctrl-c/ctrl-v.