Top 6 prompts worth saving for ChatGPT and Claude
"Where do you go to find good prompts?" someone asked on r/PromptEngineering. The same forum has a thread titled simply best system prompt for ChatGPT, and on Hacker News people have asked what custom instructions do you use for ChatGPT and is there a good collection of writing prompts.
The honest answer is that most prompt lists are long and unsorted. This one is six prompts, each with a clear job. All six come from the public prompts.chat collection and carry a CC0 licence, so you can copy, edit and reuse them without asking anyone. They are plain text: they work in ChatGPT, Claude or any other chat model, though we have not measured how well each one does on each model.
1. General assistant system prompt
General assistant system prompt (CC0, by Hector Verdugo) sets the model up as a general helper for admin and information tasks across many areas.
Good for: a starting point for ChatGPT custom instructions or a Claude project's instructions, if you have never written one.
The catch: it is broad on purpose. Once you know what you mostly ask for, add two or three lines about your own work; a generic system prompt gives generic answers.
2. GPT-5 expert prompt engineer mode (condensed)
GPT-5 expert prompt engineer mode (CC0, by michael.727) asks the model to act as a prompt engineering expert and give explanations and practical examples.
Good for: turning a vague request into a structured prompt before you run it.
The catch: the name says GPT-5, but it is ordinary text and runs on other models too. How well it does there is not measured.
3. Hallucination vulnerability prompt checker
Hallucination vulnerability prompt checker (CC0, by Scott Malin) reads one of your prompts and points out the places that invite the model to make things up or assume too much.
Good for: anyone writing prompts that feed reports, customer replies or anything where a made-up fact costs you.
The catch: it checks the prompt, not the answer. It will not tell you whether a specific claim in the output is true; you still verify that yourself.
4. Information gathering prompt
Information gathering prompt (CC0, by Mikimano) builds a research outline for any subject you want to write about: basic information split into sub-categories, or deeper specialist information.
Good for: the research step before an article, a report or a study plan.
The catch: it produces a lot of structure, so the chat runs longer. Trim the categories you do not need before you ask for the next step.
5. Design system extraction prompt kit
Design system extraction prompt kit (CC0, by Nureddin Gokbey Inac) tells Claude to go through a whole codebase and pull every design token, pattern and component into a structured JSON inventory.
Good for: cleaning up a product whose colours, spacing and buttons have drifted over time.
The catch: the model has to be able to read your code, so it fits Claude Code or a tool with file access better than a plain chat window. The output is a raw list, not a finished design system.
6. Kickstart prompt for web UX and UI design
Kickstart prompt for web UX & UI design (CC0, by Nureddin Gokbey Inac) is the author's own starting prompt for every web design project.
Good for: opening a new site or landing page with the right questions about users, goals and structure.
The catch: it gives you a UX and UI frame, not code. You still need a separate step, or a coding agent, to build it.
How to keep the good ones
The reply most of those threads end with is the same: keep your own file. Save the prompts that worked, note which model you ran them on, and edit them as you learn. A practical order:
- Start with the system prompt (1) and add a few lines about yourself.
- Run new prompts through the prompt engineer (2), then the hallucination checker (3) before anything important.
- Keep the task prompts (4, 5, 6) next to the work they belong to.
Each item page has the full prompt text to copy. Read it once before you paste: a prompt is just instructions, and you should know what you are telling the model.