My AI prompt libraries, and which one to use

I started building AI assistants by making custom GPTs, and I never really stopped. Somewhere along the way I accumulated a lot of prompts (system prompts mostly, plus a long tail of snippets, role definitions and one-off tests), and because I like to open-source what I can, they ended up scattered across a dozen public repositories and a couple of Hugging Face pages.

This post is the map. If you landed here looking for a system prompt library, or for prompts for a specific tool (NotebookLM, Open WebUI, Roo Code, ChatGPT), here is what exists, what is in each one and which to open first. Everything below is public and free to reuse.

Start here: the big system prompt library

System-Prompt-Library is the main one. It is an index of system prompts for AI agents, assistants and workflows that I have been building since the summer of 2024. It is close to a copy of my own working library, with very few entries excluded.

It is not hand-maintained. I record each assistant in a NocoDB table through a form, with a flag for whether it is OK to open-source (almost all are). An n8n workflow checks the flag, passes the JSON to an agent that writes a more readable model card, and commits the result to the repository. Periodic snapshots are exported as releases. The README counted 1,290 prompts when it was last regenerated on 28 September 2025, and it publishes the index as Markdown, CSV and JSON so you can load it into whatever you like.

Two things to know before you rely on it. There is some duplication, because I sometimes rewrite a prompt rather than version it, and some apparent duplicates are really earlier iterations of the same idea. And I replace personal details ("the user, Daniel") with generic wording by script, which is good enough for most prompts but not perfect.

The same collection is on Hugging Face as the dataset danielrosehill/System-Prompt-Library (CC BY 4.0, a little over 800 downloads at the time of writing). The dataset card describes 923 prompts in JSON and 851 in Markdown, so it is an older snapshot than the GitHub count above. Use the dataset if you want to load it programmatically, and the GitHub repository if you want the latest.

If you only want writing help, Writing-System-Prompts is a smaller dataset cut from the big library: prompts for editing, proofing and reformatting text, plus email utilities, biographies and a few for-fun ones such as an irritable editor. It has a browsable index with a one-line description per prompt.

Prompts for specific tools

These are the ones that tend to get found from search, because they answer a concrete question.

• NotebookLM-Custom-Prompts is my most-starred prompt repository. NotebookLM is polite and deferential, which is not what I want when I am working out whether a 60-page industry report contains one original idea. These are blunter prompts you paste into its custom prompt box (or use anywhere else): a sceptical "jaded report reader" persona, quick checks such as "worth the read?" and "original thinking?", and extraction prompts for the main arguments.

• OpenWebUI-Prompt-Library holds prompts written for the Open WebUI slash-command menu. They are not conversation starters. They steer a conversation that is already running: change the output format, rewrite the last answer, summarise for a hand-over to another agent, set context quickly.

• RooCode-Mode-Prompts is my set of role and mode prompts for Roo Code. My working model is that a role prompt defines what kind of team member the agent is, and a mode prompt tunes what it is doing at a given moment. I keep a few roles and many narrow modes. The repository includes JSON exports and scripts for exporting them.

• ChatGPT-Custom-Instructions is less a library than a guide to how custom instructions work: they are a system prompt wedged between the model and every turn of the conversation, so what you put there is paid for on every message.

• Speech-To-Text-System-Prompt-Library is a small skeleton of short prompts for cleaning up raw dictation with an LLM, designed to be stacked: a basic cleanup layer, then a format layer (business correspondence, say) on top. If you dictate a lot, this is the idea worth borrowing.

Prompts for writing prompts

If you would rather generate your own system prompt than copy mine, two repositories and one Hugging Face Space cover that.

System-Prompt-Generation-Configurations is a set of system prompts for assistants whose job is to write, edit and improve other assistants' system prompts. The one I use most "forks" an existing prompt: take this prompt for X and adapt it for Y. I prefer that to batch-generating prompts from nothing. Another generates two versions in one go, one for my own use and a redacted one suitable for open-sourcing.

Structured-System-Prompt-Generator is a small Streamlit app. You paste a system prompt and it returns a tidied version, a table of the data fields the prompt needs and a JSON schema for structured output. It calls OpenAI's API, so you need your own key.

The Hugging Face Space System-Prompt-Factory is a modular Streamlit utility for the same job. When I checked on 3 November 2026 it was showing a runtime error, so treat it as not working until I have looked at it.

General-purpose and reference collections

General-System-Prompts tackles an odd problem. With no system prompt, a model feels flat. With a detailed one, it becomes an assistant. I wanted a light-touch prompt that gives a neutral, general-purpose feel, like the experience in ChatGPT. The repository holds a few of mine, personality definitions, and system prompts published by the major vendors with brand references stripped, as templates.

Prompt-Library is the oldest of the group (last touched December 2024): prompts I have used with example outputs and notes. Its most useful idea is "context snippets", small pieces of text about you or your work that you append to prompts. RAG is far more powerful, but a folder of snippets is something you can start today.

Text-To-Image-Test-Prompts is the odd one out, since it is for image models. It has 32 categories of roughly five test prompts each (typo handling, lighting instructions, logo design, character consistency, confusing or conflicting instructions and so on), meant for people who want to try a model on their own use case rather than read a benchmark.

Awesome-LLM-Prompt-Libraries is an index of other people's open-source prompt libraries (not my own prompts), grouped by general, ChatGPT, specific field (cybersecurity, data science, game development and more) and non-English. Tutorials and courses are out of scope. It has not been updated since November 2024, so some links may have gone stale.

Which one should you use?

• You want a large set of ready-made system prompts to search: System-Prompt-Library (GitHub for the latest, Hugging Face for loading in code).

• You only need writing and editing prompts: Writing-System-Prompts.

• You read a lot of long reports: NotebookLM-Custom-Prompts.

• You use Open WebUI: OpenWebUI-Prompt-Library.

• You use Roo Code: RooCode-Mode-Prompts.

• You dictate and want clean text out: Speech-To-Text-System-Prompt-Library.

• You want to write your own system prompts: System-Prompt-Generation-Configurations, or the Structured-System-Prompt-Generator app.

• You want other people's prompt libraries: Awesome-LLM-Prompt-Libraries.

What I would do differently

None of these was empty, but they are uneven. The big library is a database with a front end; the small repositories are individual experiments and some have not moved in a year or more. The Space needs fixing. And I have a habit of starting a new repository when a folder in an existing one would have done, which is how I got to twelve.

The consolidation I would like is one library with the tool-specific collections as folders. Until then, this page is the index.