Sources and licences
Every prompt links to the exact upstream file or post, pinned to the commit we read. Where the licence allows reuse (CC0, MIT, Apache-2.0, CC BY-NC-SA) we quote the full text with credit. Where there's no licence, or the licence attaches conditions to copies, we show only a short excerpt with our own summary and link to the original.
| Source | Prompts | Author | Licence | How we use it |
|---|---|---|---|---|
| Arnie936 / llm-prompting-tests | 22 | Arnie (GitHub @Arnie936) | No license file (all rights reserved) | The quality bar for this library. There's no license, so we show a short excerpt, our own summary and grading notes, and link to the exact prompt. |
| KCORES LLM Arena | 4 | karminski (KCORES team) | KCORES License 1.0 (custom: copies must carry the full licence text and public source) | Each test comes with a published point-by-point scoring rubric. The licence attaches conditions to copies, so we link to the full prompt and summarise the rubric in our own words. |
| KCORES prompt collection (gist) | 3 | karminski-牙医 | CC BY-NC-SA 4.0 | Quoted in full with attribution for non-commercial use. The English translations are ours (BranoLabs), marked as translations and shared under the same licence. |
| LocalLLM_VisualCodeTest | 3 | David Harris (GitHub @electricazimuth) | No license file (all rights reserved) | Built for testing local models through llama.cpp or KoboldCpp, with published side-by-side results. No license, so excerpt plus link. |
| flight-combat-llm-comp | 2 | Alexandru Vasile (GitHub @AlexTzk) | No license file (all rights reserved) | The same prompts run on 9 local MLX models, with the outputs published next to them. No license, so excerpt plus link. |
| Misguided Attention | 16 | GitHub @cpldcpu and contributors | CC0 1.0 | Public-domain dedication, quoted in full. Only prompts from the public README: the author keeps the longer eval set scrambled to limit training contamination, and we respect that. |
| SimpleBench (public sample) | 4 | SimpleBench team | MIT | Quoted in full under MIT with the official answer key. Only 10 questions are public; the real benchmark is private. |
| IFEval (Instruction-Following Eval) | 4 | Google Research (Zhou et al., 2023) | Apache-2.0 | Quoted in full under Apache-2.0. Every constraint is machine-checkable; this site's checker runs similar checks in your browser. |
| Berkeley Function Calling Leaderboard (BFCL) | 2 | Gorilla team, UC Berkeley | Apache-2.0 | Question, function schemas and ground truth from the BFCL data files under Apache-2.0. Our change: wrapped into one copy-paste prompt for chat apps without a tool API. |
| Pelican on a bicycle | 1 | Simon Willison | No explicit license; one-sentence prompt quoted as a citation | A single sentence, quoted with attribution. |
| Spinning hexagon (X post) | 1 | Flavio Adamo (@flavioAd) | No explicit license; two-sentence prompt quoted as a citation | The viral o3-mini vs DeepSeek R1 test from 31 Jan 2025, quoted with attribution. |
Looked at and left out
Quality over quantity: these didn't make it, and here's why.
| Source or prompt | Why not |
|---|---|
| KCORES Snake Pit test | Unfinished: its scoring table is a copy of the solar-system rubric, so there is no valid way to grade it. |
| AlexTzk bug-hunt task | Good idea, but there's no published list of the planted bugs, so we can't give a verified answer key. |
| Misguided Attention long eval set (52 prompts) | The author scrambles it on purpose to keep it out of training data. Publishing it in clear text would undo that. |
| Misguided Attention: Knights and Knaves (Easy) | Two assignments are consistent (A knight and B knave, or the reverse), so 'determine who is who' has no single answer. |
| Misguided Attention: poisoned hot tea, Not from Hanoi | No single correct answer, or wording too ambiguous to grade fairly. |
| BFCL parallel_multiple_3 (rectangle with perimeter 14 and area 15) | No real rectangle has those values (l + w = 7 and l × w = 15 has no real solution), yet the ground truth expects plain calls. A careful model gets marked wrong. |
| EQ-Bench Creative Writing v3 prompts | No licence file, and grading relies on a judge model rather than criteria you can check yourself. |
| AidanBench | No licence; open-ended questions scored by embeddings and a judge model, with no expected answers. |
| lechmazur confabulations and NYT Connections benchmarks | No licence, and they need their document sets or harness; they aren't copy-paste prompts. |
| Needle in a haystack (gkamradt) | Licence not stated, and it needs generated long documents rather than a single prompt. |
| SnitchBench | MIT and interesting, but it needs a tool-calling harness and multi-turn runs. A candidate once we add harness-based tests. |
| scka-de/vibe-coding-art, magiruuvelvet/llm-prompts | System prompts for generating art or for daily use, not tests with grading criteria. |
| Generic 'write Snake / Tetris / a todo app' prompts | So common in training data that almost every model passes; they no longer tell models apart. |
| electricazimuth heptagon, chicken, fireworks and light-show prompts | Covered by stronger versions already in the library (KCORES heptagon) or less discriminating than the three we kept. |
Are you an author and want a prompt changed, credited differently or removed? Use the bug link in the footer and we'll fix it quickly.