Every AI product reads a block of instructions before it ever sees your message. Who it is, what it refuses, how it formats answers, what it must never say. You do not write it and you are not shown it.
One repo has captured 430 of them, organised by vendor. The leak angle is the boring part. Here is what is actually useful.
The 13x gap
Anthropic publishes its Claude system prompts. Openly, on purpose, in the release notes. The repo stores those too, so you can put the published version next to the captured one.
They are not the same size:
- Official, published by Anthropic: 4,269 words
- Captured from the running product: 56,148 words
Roughly 13 times larger.
That gap is not a scandal. Published prompts legitimately exclude tool definitions, per-surface instructions and operational scaffolding. But if you were studying the published version to learn how a frontier team writes prompts, you were reading a summary and did not know it.
Dropping both into a text diff is the fastest way to see what the published version leaves out. That is the exercise I would actually recommend from this repo.
What the prompts reveal
Read enough and patterns show up that no marketing page would state.
Enormous effort goes into tone, not capability. Large fractions of every prompt are spent on how to sound. Do not open with flattery. Do not pad with caveats. Do not moralise. Vary sentence length. Providers clearly treat default LLM voice as a serious product problem and spend real prompt budget on it.
Refusals are far more specific than public policy. Published safety policies are broad and principled. The prompts contain concrete enumerated categories. If a model ever refused something oddly specific, this is where that lives.
If you write system prompts for your own product, this is the closest thing to watching how the best-funded teams in the industry solve your exact problem. The formatting rules alone will improve your output.
Stay sceptical
The repo says prompts are captured verbatim. Treat that carefully.
These are extracted by asking the model to repeat its instructions. It works because the prompt genuinely is in context and the model can read it back. But a model reproducing text from its context is not a guaranteed faithful copy. It can paraphrase, skip, smooth a section, or in the worst case confabulate something plausible.
Treat them as high quality evidence, not primary source documents. Patterns across many files are trustworthy. Any single exact sentence is less so.
Longer version covering what a system prompt is from scratch, the full vendor breakdown, and whether any of this is legal: What is a system prompt? 430 real ones, captured from every major AI.
Top comments (0)