Skip to content

CONTRIBUTING.md

Latest commit

 

History

History
83 lines (51 loc) · 4 KB

CONTRIBUTING.md

File metadata and controls

83 lines (51 loc) · 4 KB
Raw

Contributing

Thanks for contributing to The Big Prompt Library. The rules below exist so the collection stays trustworthy and navigable.

The one rule that matters: verbatim only

Contribute the prompt exactly as the system emitted it. Do not summarize, paraphrase, condense into bullets, clean up, or translate.

A summary is not useful here. The value of this repository is that a reader can see precisely how a real product prompts its model - the wording, the ordering, the hedges, the typos. A condensed version destroys all of that and cannot be verified by anyone.

Long is fine. Extracted coding-agent prompts in this repo routinely run 500+ lines.

If you can only capture part of a prompt, say so explicitly and label the entry Partial.

Where things go

Directory Contents
SystemPrompts/<Vendor>/ System prompts from LLM products and services
CustomInstructions/ChatGPT/ Custom GPT instructions
Jailbreak/<Vendor>/ Jailbreak prompts
Security/GPT-Protections/ Instruction-protection techniques
Articles/<topic>/ Write-ups and analysis

Naming

Use one of these two forms:

  • YYYYMMDD-ProviderName.md - e.g. 20251202-Claude-Opus4.5-Soul-Document.md
  • provider_feature_YYYYMMDD.md - e.g. github_copilot_cli_20260221.md

Custom GPTs are named {gpt_id}.md; generate the file with python Tools/openai_gpts/idxtool.py --template <CHATGPT_URL>.

Provenance header

Every new entry starts with a header block so readers can judge and re-verify it:

# <Name> (unverified)

- **Contributed by**: [username](https://github.com/username)
- **Contributed on**: MM/DD/YYYY
- **Source**: <product, tier, platform - e.g. "Android Gemini App (Paid tier)">
- **Note**: <extraction method, discrepancies, anything partial or uncertain>

Drop (unverified) only if the prompt is officially published by the vendor or you can point at a public source. A link to a gist, tweet, or vendor repo in the Source line is worth more than any assertion.

If you are not the original extractor, credit whoever was.

Update the index in the same pull request

  • SystemPrompts/README.md is maintained by hand. Add your entry to the right vendor section - newest first. If the vendor has no section yet, add one before ## Miscellanous, with a See: <vendor url> line.
  • CustomInstructions/README.md is generated. Run python Tools/openai_gpts/idxtool.py --toc and commit the result.
  • Jailbreak/, Security/, and Articles/ each have their own README.md to update.

A file nobody can find from an index may as well not be in the repo.

Multi-file agent contexts

Most contributions are a single markdown file. When a system exposes its context as many files - a module tree, a set of AGENTS.md files, tool declarations - keep the structure and add a folder instead:

SystemPrompts/<Vendor>/<provider_feature_YYYYMMDD>/
    README.md      <- provenance header, what's captured, what's missing, an index
    ...            <- the files, unmodified

Link the folder's README.md from SystemPrompts/README.md. See SystemPrompts/Notion/notion-ai_20260322/ for a worked example.

Source files are in scope when they are what the model sees - tool and type declarations, instruction files. A product's actual implementation source is not.

No secrets, no other people's data

Sandbox and workspace dumps leak. Before opening a pull request, check your files for API keys, tokens, session identifiers, real email addresses, and any workspace content belonging to real people. Redact it, and note the redaction in the entry.

Screenshots

Evidence screenshots are welcome next to the prompt they support (see SystemPrompts/xAI/08212024-Grok2-fun.png). Keep them small and crop out anything personal.

Scope

This repository is for educational and research use - understanding how these systems are prompted, and how prompts leak. Contributions aimed at operationalizing abuse aren't accepted.