Skip to content
HN On Hacker News ↗

One Shot Challenge — One Prompt, One AI Model, One Shot

▲ 11 points 10 comments by lucasmartinic 1mo ago HN discussion ↗

Pangram verdict · v3.3

We believe that this document is a mix of AI-generated, and human-written content

65 %

AI likelihood · overall

Mixed
29% human-written 71% AI-generated
SEGMENTS · HUMAN 1 of 4
SEGMENTS · AI 3 of 4
WORD COUNT 1,143
PEAK AI % 99% · §4
Analyzed
Jul 13
backend: pangram/v3.3
Segments scanned
4 windows
avg 286 words each
Distribution
29 / 71%
human / AI fraction
Verdict
Mixed
Pangram v3.3

Article text · 1,143 words · 4 segments analyzed

Human AI-generated
§1 AI · 76%

Boil down your creativity and vision into your best shot.Round 01Jul 13 – Jul 18, 2026 · winners Jul 196 days. 6 winners. 6 GTA VI copies, plus a PlayStation 5 for first place.One prompt in, something spectacular out. Build a whole game, a living world, a toy that feels impossible. The more complex, rich and deep, the better.A year ago this was impossible.The newest models, like Fable 5, Opus 4.8, GPT-4.6 Sol, Gemini 3.1 Pro, Grok 4.5, GLM 5.2, DeepSeek V4, Seed 2.1 Pro, Qwen3.7 Max, Kimi K2.6, Muse Spark 1.1build entire apps, games and worlds from a single prompt. No scaffolding. No back and forth. You describe it once, and a working thing comes out.The One Shot Challenge puts that to the test: how far can you and your model get in a single shot?01The prizes1st placePlayStation 5All 6 winnersGTA VIAnd a spotlightSIGGRAPH02How it works01Write one promptOne prompt, as long as you want. It is your whole entry.02Bring your best modelAny provider, your own key. No key? Muse Spark ($20 free) and Gemini give you a frontier model; Groq and NVIDIA hand out free keys too.03Confirm your emailClick the link we send. Nothing runs until you do.04It runs onceOne model call. No edits, no retries. Broken still counts.05It ships, sealedInto the gallery, prompt and model hidden until results.066 winnersFive picked by a jury (Lionsgate AI, Sony, the NBA, World Labs), one by your community vote.03Under the hoodNo black box. We run your prompt on our own server for one reason: so every entry is a provable single prompt and single model call, with no edits or retries. Your API key is encrypted, used for exactly one call, then destroyed, never stored with your entry.

§2 Human · 10%

The result is published in a locked-down sandbox with a hash anyone can verify. Every entry is untrusted code, so it runs walled off and isolated: it can't reach our servers, your data, or any other entry.See the full architecture and security →04The rulesRuns July 13–18. One entry per day, per email: 6 shots total.Prompts and models stay sealed until the results.A jury of AI leaders from Lionsgate AI, Sony, the NBA and World Labs picks 5 of the 6 winners at the Worlds in Action Hack at SIGGRAPH LA on July 19. The 6th is the community favorite, decided by your votes.Anything goes. 3D worlds, games, toys, sandboxes and experiments are encouraged.Your entry must be the self-contained output of your one prompt. Libraries (three.js and the like) from the approved CDNs are fine, but loading your own external or mutable code bundle, so you can keep editing it after submission, is grounds for disqualification.Use only assets, code and fonts you are licensed to use. Free and open sources, like the CC0 textures and models on Poly Haven, are perfect. Here’s how to load assets from the CDNs.Submissions close in5days13hrs16min41secThe challenge, live0shots fired0tokens used0.0hrsmodels have worked05Latest entriesSee all 14 →06QuestionsIs it free to compete?Yes. Entry is free and there is no fee to us, ever. The only cost is your own model run on your own API key (BYOK). Bring your best model and go. If you’d rather not pay for tokens, there are solid free options too. Meta’s Muse Sparkgives every new account $20 in free credits, plenty for many one-shot runs on a strong frontier coding model (US-only for now). Anywhere else, use Google Gemini’s free tier, a frontier model too (rate-limited, but plenty for one shot a day). Or grab a free key from Groqor NVIDIA, which run open-source models, so tougher odds but a great prompt can still win.Enter my API key? Are you CRAZY!?Fair reaction.

§3 AI · 91%

Here is the deal. Your key is encrypted the moment it arrives, used for exactly one model call, then permanently deleted. It is never stored in plain text, never written to logs, and never shown in any page or API response. Tip: create a fresh key with a spend limit and revoke it right after you submit.Why run it on our side at all, instead of letting you paste in a result? Integrity. This is a one prompt challenge, so every entry has to be a provable single prompt and single model call, with no edits, retries, or chaining. Running it on our server with your key is what guarantees that, and lets us publish a hash of your prompt plus the exact output so anyone can verify it. If everyone ran it themselves and pasted results, there would be no way to keep it fair. And our server is locked down: no open public ports, reachable only through a Cloudflare Tunnel.How are entries judged?The judging criteria aren’t completely defined yet, but creativity, depth, aesthetics, and functionality are some of the things that will be considered. A jury of AI leaders picks 5 of the 6 winners; the 6th is the community favorite, decided by your likes.Can I use images, 3D models, or fonts?Yes. Your entry runs in a locked-down sandbox, but it can pull libraries, fonts, and open assets from an allowlist, including CC0 textures and models from Poly Haven. The trick: the model can’t browse, so paste the asset’s exact URL into your prompt. Full guide with examples: how to load assets from the CDNs.Do you add anything to my prompt?One system message, and it is the same for everyone. It only tells the model the output format: a single self-contained HTML file that runs in our sandbox. No creative or quality hints, so the ideas are all yours. You can read the exact text right on the submit form and on the architecture page.Will there be more rounds?Yes. This is the first round, and every round after brings bigger prizes. Leave your email and we'll tell you the moment the next one opens.What can I build?One self-contained HTML file. You can pull in libraries like three.js from approved CDNs, but there is no build step and no backend. Think games, toys, simulations, generative art, 3D worlds.

§4 AI · 99%

Anything that runs in a single page.Why only one file, no backend?So every entry is instantly runnable, easy to verify, and safe to sandbox in an isolated frame. One file in, one file out. It also keeps the field even: everyone ships the same shape of thing.What does “one shot” actually mean?Your prompt is sent to the model a single time. No edits, no retries, no follow-up messages. Whatever it returns, polished or broken, is your entry. You get one shot per day.When are the prompts and models revealed?Not until the results. During the round every entry's prompt and model stay sealed, so nobody copies and the model stays a mystery. Each entry still ships with a hash of the prompt and output, so anyone can verify it really was one shot.Are you and your model a tight team?6 days. One shot a day. Free to enter.Prove yourself