Ruocco Academy
RUOCCO ACADEMY

Every prompt. Every line of code. Behind every video.

Join the waiting list The Academy is not open yet: leave your email and you get one message when it opens.

In my videos you see AI models build games and apps live. In the Academy you get exactly how: the prompts I used, word for word, the settings, the step-by-step instructions, and the full source code of every demo, ready to download. Plus videos only members can watch.

52video projects
224exact prompts
513demos with full code
1members' videos

What's inside

The prompts

The exact words each model got, with model name, reasoning effort, tokens, time and cost. Copy them and run them yourself.

The instructions

How each demo was made and tested: the setup, what broke, what I fixed by hand and why. No guessing.

The code

Every game and app as a single file you can open, study and remix, plus a ZIP of the whole project.

Projects in the library

GPT-6.2 Astra vs GPT-6.2 Sol vs Claude Fable 5.2: seven prompts, the real bill

OpenAI's GPT-6.2 Astra (November 17, 2026) costs the same per token as Claude Fable 5.2 ($8 / $40). Astra and its cheap sibling Sol got seven prompts (landing page, 12-slide deck, trip planner, finance dashboard, kids' game, night-harbour shooter, mountain lake), Fable the two 3D ones. API-equivalent bill: Astra $8.29 for 7 builds, Sol $1.95, Fable $20.48 for 2. Every prompt, the runner and every file as the models wrote it are here.

7 prompts16 demos ๐Ÿ”’ members

One episode in, a month of content out: the content pack instructions + the automation

The repurposing service from the video: Claude Opus 5.5 turns one episode transcript (with [mm:ss] times) into 5 Shorts with exact cut points, a newsletter, 5 LinkedIn posts, an X thread, YouTube titles and chapters and a 4-week calendar, in one page with copy buttons. Here: the exact instructions, the runner, the 35-line folder automation and the script that cuts a Short from Claude's times.

2 prompts2 demos ๐Ÿ”’ members

Qwen 4 Max vs Qwen 3.8 Max: 60% cheaper? The real bill on the same six prompts

Alibaba's Qwen 4 Max (November 16, 2026) is said to do Qwen 3.8 Max's work for about 60% less. Both got, word for word, the six prompts Opus 5.5, Sol and three Gemini models already built. Qwen 4 Max: 6 of 6 builds running as written, $0.89 at list price (196k output tokens). Qwen 3.8 Max: 2 of 6, $2.05 (341k). 57% less on our bill. No edits on any Qwen file.

6 prompts24 demos ๐Ÿ”’ members

Gemini 4 Carbon vs Argon, Hydrogen, Opus 5.5 and Sol: same six prompts, the real bill

Google's strongest model, Gemini 4 Carbon (November 10, 2026), costs exactly what Argon costs per token. Given word for word the six prompts Argon, Hydrogen, Opus 5.5 and Sol already built, it made the best Gemini builds yet, but wrote 482k output tokens against Argon's 152k: $5.21 at list price against $1.52, and $10.43 from December 1, 2026. Includes the route cap, the continuation requests and three marked one-line edits.

6 prompts33 demos ๐Ÿ”’ members

Gemini 4 Hydrogen vs Argon vs Claude Opus 5.5 vs GPT-6.1 Sol: same six prompts, the real bill

Google's new fast model, Gemini 4 Hydrogen (November 12, 2026), given word for word the six prompts Argon, Opus 5.5 and Sol already built: a life animation, a voxel castle school, a short film, a 3D car showroom, a driving game and a music studio. On high: 6 of 6 run, one needed a one-line fix to show any picture, $1.75. A bonus run on xhigh: every request stopped at the 65,536-token output limit of its route ($2.95, 3 complete files).

6 prompts28 demos ๐Ÿ”’ members

Claude Fable 5.2 on the $20 plan vs Muse Ember vs a free model: one FPS, one lake

Two prompts we wrote, a small night-harbour shooter and a mountain lake at golden hour, sent once to Claude Fable 5.2 on the Claude Pro plan, Meta Muse Ember on OpenRouter and MiniMax M3.1 Flash free in OpenCode. The exact prompts, the runners and every file as the models wrote it are here.

5 prompts6 demos ๐Ÿ”’ members

Three viral-format 3D Shorts written as code by Claude Opus 5.5 (three.js + MakeHuman)

"What If ___ Disappeared For 5 Seconds?": glass, oxygen and your skeleton. Claude Opus 5.5 on the Claude Pro plan wrote the scripts, every three.js animation (timed to the voice to the hundredth of a second) and a setup script that downloads portable Blender, the MakeHuman plugin and its CC0 assets and builds a rigged 3D human. All ten prompts word for word, including every fix, are here.

10 prompts3 demos ๐Ÿ”’ members

Ripple-65B and Hy4 on the exact prompts Kimi, Opus 5.5 and GPT-6.1 Sol got, plus our weight-space check on an RTX 3060

Ricursive's Ripple-65B Preview and Tencent's Hy4 (both released November 4, 2026) given, word for word, the six prompts of the Opus 5.5 vs GPT-6.1 Sol test (life animation, voxel wizard castle, short film, SUV showroom, coastal driving game, music studio). Hy4: 6 of 6 run ($0.23). Ripple: 0 of 6 run as written ($0.17). Plus the script and raw results of our MATH-500 check of the MIT weight-space post on Qwen3.5-0.8B.

6 prompts33 demos ๐Ÿ”’ members

Herdr full test: Claude Code directs Codex, Sonnet, Gemini and Grok, and talks to an agent on another machine

Inside Herdr, Claude Opus 5.5 opened a 2x2 grid, started Codex, Claude Sonnet 5.5, Gemini 4 Argon and Grok 4.7 and gave each one scene of an ASCII film, then played them together. Here are every prompt, TEAM.md, the four scene programs as the agents wrote them, the permissions used and the remote setup.

6 prompts5 demos ๐Ÿ”’ members

Claude Code with a team vs alone, in Herdr: two races

Claude Opus 5.5 led Codex (GPT-6.1 Sol) and Claude Sonnet 5.5 through Herdr, against the same Opus working alone. Same brief, same second, both recorded. Here are the exact prompts, both briefs, the contract the lead wrote, the permissions used and the four builds as the agents left them.

4 prompts4 demos ๐Ÿ”’ members

Three real-life Claude ideas, one prompt each: room planner, drone lab, wall panel

Ideas people are using Claude for, given once each to Claude Opus 5.5 on the Claude Pro plan with every tool off: a 3D room planner (will the new bed fit?), a drone lab you build part by part and fly, and a smart-home wall panel for an old tablet. The exact prompts, the runner that sent them and the three files as the model wrote them are here.

3 prompts3 demos ๐Ÿ”’ members

GPT-6 Astra vs Claude Opus 5.5: write a neural network's weights by hand

One prompt asked each model to write all 1,210 weights of a 64-16-10 digit reader, with no examples, no code and no tools. The exact prompt, the scripts that sent it and scored it on 1,797 real digits, both full answers and the Qwen3.5-0.8B loop test are here.

1 prompts0 demos ๐Ÿ”’ members

Impeccable vs Hallmark vs no skill: the same prompt to Claude Opus 5.5

Two design skills for coding agents on one brief, a one-page site for a fictional bike shop in Naples, plus a run with no skill. The exact prompts, the command that ran each build and the three pages as the model wrote them are here.

3 prompts3 demos ๐Ÿ”’ members

LFM3 1.7B vs Spark-X 1.7B on an RTX 3060, with the exact commands

Two 2B open models on a 12 GB RTX 3060 with llama.cpp: the fact question, the JSON extraction, three Pong runs each and three prompts from the Opus 5.5 / Sol test, word for word. Exact commands, prompts and scripts are here, and every file is as the models wrote it.

4 prompts12 demos ๐Ÿ”’ members

The free 32 GB Kaggle GPU vs my RTX 3060, with every notebook cell

Qwen3.8 27B on Kaggle's free T4 x2 and on a 12 GB RTX 3060, on two prompts from the Opus 5.5 / Sol test, word for word. Every notebook cell, the tunnel tests, the speed tests and the marked fixes are here.

2 prompts7 demos ๐Ÿ”’ members

Seven AI video models on one prompt, with the checklist and the real bill

Kling 4.0, Seedance 2.5, MiniMax H3, Wan 3.0 Prime, Grok Imagine 1.5, FLUX 3 Video and Veo 3.1 on one prompt with eight checkable traps. One attempt each. The exact prompt, the API settings for every model, the grading and the cost of each clip are here.

1 prompts0 demos ๐Ÿ”’ members

Build Your Own LLM From Scratch, Ep. 4: how a network learns (backpropagation by hand, then PyTorch)

Everything from episode 4: three neurons solve the staircase light, first with knobs set by hand, then by themselves with backpropagation written in plain Python, and finally the same network in about twenty lines of PyTorch. With the fix for the Visual C++ error you may hit on a clean Windows, and the exact output of the video.

0 prompts2 demos ๐Ÿ”’ members

Does more thinking pay? Claude Opus 5.5 at effort low, high and max vs Gemini 4 Argon, with the real bill

Three builds of the kind posted for Gemini 4 Carbon (a robotic eye that explodes into its parts, a mechanical flower, a luxury kitchen), given word for word to Claude Opus 5.5 at reasoning effort low, high and max and to Gemini 4 Argon at effort high. One request each through OpenRouter. The three max runs thought until the 128k output limit and returned no code; every time, token count and cost is here.

12 prompts12 demos ๐Ÿ”’ members

Every way to 128 GB vs a 12 GB RTX 3060: Qwen3.8 Flash Next local on the Opus 5.5 / GPT-6.1 Sol prompts

Qwen3.8-Flash-Next IQ3_S (3-bit, 78 GB) running locally in Strata 0.2.7 on an RTX 3060 12 GB + 64 GB RAM, given word for word the six prompts of the Opus 5.5 vs GPT-6.1 Sol test. 52 minutes, $0. Three builds run after small marked fixes, three do not run as written. Plus every command and measurement: llama.cpp b11954/b11960 with and without the new MoE GPU cache, and Strata's own numbers.

6 prompts23 demos ๐Ÿ”’ members

GPT-6.1 Terra vs Claude Opus 5.5 vs Claude Haiku 5.5: four builds and the real bill

OpenAI's new cheap and fast model (released October 26, 2026) against the best and the smallest Claude, on the same four prompts: a steam train flying between floating islands, an animal cell in SVG, a voxel rocket launch and a floating treasure tower. One run each through OpenRouter, every time, token count and cost, and the two marked one-line fixes Haiku needed.

12 prompts14 demos ๐Ÿ”’ members

Build Your Own LLM From Scratch, Ep. 3: your first artificial neuron

Everything from episode 3: one artificial neuron in plain Python that learns its own knobs from four examples (Rosenblatt's perceptron rule), the same code on different data, and the staircase light it cannot learn. No libraries, no randomness: exactly the output of the video.

0 prompts1 demos ๐Ÿ”’ members

Kimi K3.1 Pro on the exact prompts Opus 5.5, GPT-6.1 Sol and Fable 5.1 got: eight builds and the real bill

Moonshot's Kimi K3.1 Pro (released October 23, 2026) given, word for word, the six prompts of the Opus 5.5 vs GPT-6.1 Sol test (life animation, voxel wizard castle, short film, SUV showroom, coastal driving game, music studio) and the two prompts of the Fable 5.1 / Opus 5.5 / GLM 5.3 Flash test (KAZE BLOCKS, sky islands). Eight new builds, $1.69; the rivals' builds are the originals.

8 prompts29 demos ๐Ÿ”’ members

Qwen3.8 Flash Next (Strata) vs Qwen3.8 27B in OpenCode at full context on a 12 GB RTX 3060

A 125B mixture-of-experts model and a dense 27B, both local at 256K/262K context, as the model behind the OpenCode agent on one gaming PC (i5-12400, 64 GB RAM, RTX 3060 12 GB). The exact project prompt, the OpenCode config for Strata, the llama.cpp launch line, the measured speeds and all four builds (two Flash Next runs, Qwen original + 2-line fix).

2 prompts4 demos ๐Ÿ”’ members

Liquid LFM2.5 24B vs Qwen3.8 27B on a 12 GB RTX 3060: settings, speeds and three real jobs

Two open models run offline on one gaming PC (i5-12400, 64 GB RAM, RTX 3060 12 GB) with llama.cpp: every launch command we used, the measured speeds (Qwen 2.8 -> ~20 tok/s, Liquid ~100 tok/s, Liquid's DSpark drafter slower on this card), the document, spreadsheet and game prompts, the bench scripts and the fictional test files.

3 prompts3 demos ๐Ÿ”’ members

Recursive Flint on a 12 GB RTX 3060 vs Gemini 4 Argon vs Claude Opus 5.5: four Carbon-style prompts

The free open-weight coding model Recursive says it built with its automated research system, run locally with llama.cpp on a gaming PC, against Gemini 4 Argon and Opus 5.5 on the same four prompts (voxel pagoda, SVG controller, exploding orange, SVG soda bottle), plus the Snake game Flint wrote live. Every launch command, the measured speeds and the two marked fixes.

5 prompts14 demos ๐Ÿ”’ members

Build Your Own LLM From Scratch, Ep. 2: Python from zero to your first language model

Everything from episode 2: hello.py, the word-counting text generator (it gets stuck in a loop, then rolls loaded dice) and the same counter run on eight Sherlock Holmes books. Plain Python, no libraries, with the exact output from the video.

0 prompts3 demos ๐Ÿ”’ members

Free in OpenCode: DeepSeek V4.1 Flash-1017 vs MiniMax M3.1 Flash, three prompts

The two Flash models OpenCode serves for free (Oct 2026) got the same three prompts through OpenCode itself, one run each, $0: a night-city drift racer, the launch page of a fictional AI gadget, and a 30-second motion explainer. Six builds, two one-line fixes (both MiniMax), originals included, and the sandbox that keeps the agents in their folder.

3 prompts8 demos ๐Ÿ”’ members

Free AI motion videos: DeepSeek V4.1 Flash Free (OpenCode) inside OpenDesign Motion

OpenDesign 0.25.0's new Motion mode driven by DeepSeek V4.1 Flash Free in OpenCode, at no cost: OpenDesign's own showreel example, a 20-second vertical explainer and a new VitoMind channel intro (with one follow-up fix). Source compositions + the MP4s rendered with HyperFrames.

3 prompts3 demos ๐Ÿ”’ members

DeepSeek V4.1 Flash-1017 vs the old V4.1 Flash vs Opus 5.5 vs GPT-6.1 Sol: voxel siege, windy hill, 3D headphones

Three prompts inspired by the week's viral DeepSeek tests (a voxel kingdom under siege with a dragon, a grassy hill in the wind with sound, a 3D headphones launch page), each given once to the new DeepSeek V4.1 Flash-1017, the previous V4.1 Flash, Claude Opus 5.5 and GPT-6.1 Sol. Twelve builds, $1.49 in total.

3 prompts13 demos ๐Ÿ”’ members

The "Claude Next" Japanese voxel sandbox prompt on Fable 5.1, Opus 5.5 and GLM 5.3 Flash, plus sky islands

The short prompt behind the viral "baby AGI" clip (a Japanese voxel sandbox world in one HTML file), almost word for word, given to Claude Fable 5.1, Claude Opus 5.5 and GLM 5.3 Flash, then a VoxelBench-style floating sky islands scene. Six builds, $2.77 in total.

2 prompts6 demos ๐Ÿ”’ members

AI Preview Lab: games, music and code animation

Four original browser demos generated with Claude Opus 5.5 and Ling 3.1 Flash: an island game, sequencer, six-style animation and short film.

4 prompts4 demos ๐Ÿ”’ members

Muse Spark Open vs four open models and Muse Wonder: a survivors game, a pagoda, a city to blow up

Meta's open-weight Muse Spark Open on launch day: the same survivors-game prompt to Muse Spark Open, Inkling, HY4 Preview and Step-5 Preview (plus Nemotron 3.5 Super VL as a bonus), then Muse Spark Open against Meta's closed Muse Wonder on a voxel pagoda and a voxel city with a TNT block. Nine builds, $0.66 in total.

3 prompts9 demos ๐Ÿ”’ members

Free OpenCode models: Cat Jam vs MiMo vs Big Pickle, three prompts

Three models that are free right now in OpenCode (the new stealth Cat Jam, MiMo-V2.6-Flash and Big Pickle) got the same three prompts through OpenCode itself: a pixel-art animation studio, a 3D glider game over a sunset canyon, and the exact one-sentence civilization challenge that produced the Fable 5.5 film. Nine builds, and the sandbox that kept the agents in their folders.

3 prompts9 demos ๐Ÿ”’ members

Free AI models vs Claude Haiku 5.5: six builds, same prompts

Three free OpenRouter models (Dots3-Note Preview, Laguna S 2.1, Nemotron 3.5 Lightning) and Claude Haiku 5.5 as a paid reference got the same six short prompts, word for word: a life animation, a voxel castle school, a short film, a 3D car showroom, a driving game and a music studio. Twenty-four single-file demos, with the one repair attempt each broken file was given.

12 prompts24 demos ๐Ÿ”’ members

Gemini 4 Argon vs Claude Opus 5.5 vs GPT-6.1 Sol: six builds, same prompts

Google's Gemini 4 Argon is available to everyone. The same six short prompts, word for word, to Argon, Claude Opus 5.5 and GPT-6.1 Sol: a life animation, a voxel castle school, a short film, a 3D car showroom, a driving game and a music studio. Eighteen single-file demos.

6 prompts18 demos ๐Ÿ”’ members

Opus 5.5 vs GPT-6.1 Sol: a brick breaker, a bullet-hell shooter and a voxel giant

The same three plain behavior descriptions, word for word, to Claude Opus 5.5 and GPT-6.1 Sol: a brick breaker whose sound pans with the ball, a bullet-hell shooter with rings of bullets and a voxel giant game. Six single-file demos, each with an autopilot.

3 prompts6 demos ๐Ÿ”’ members

Opus 5.5 vs GPT-6.1 Sol: can code replace an image model? Six builds, same prompts

The animations shared for the rumoured new Opus were drawn entirely in JavaScript. The same six short prompts, word for word, to Claude Opus 5.5 and GPT-6.1 Sol: a life animation, a voxel castle school, a short film, a 3D car showroom, a driving game and a music studio. Twelve single-file demos, plus the raw VoxelBench identification check.

6 prompts12 demos ๐Ÿ”’ members

Fable 5.1 vs Kimi K3: four builds, same prompts

How close is the best Chinese open model to the best you can buy? The same four prompts, word for word, to Claude Fable 5.1 and Kimi K3: a game controller drawn in SVG, a night-time tactical shooter, a roguelike with real lighting and a spreadsheet app. Eight single-file demos.

4 prompts8 demos ๐Ÿ”’ members

Gemini 3.8 Flash vs Claude Opus 5.5: three builds, same prompts

Is Google's Flash really at Opus level? The same three prompts, word for word, to both models: a floatplane simulator (the test used on the leaked Gemini 4 checkpoints), a village of NPCs that remember you and a swarm of bug-hunting agents. Six single-file demos.

3 prompts6 demos ๐Ÿ”’ members

Ube Latte Cafe: Google Playground's launch prompt, built in three chat messages

Google's Playground (US-only) builds games from a sentence. The prompt from its launch video, a cozy ube latte game in a busy retro cafe, given to Gemini 3.8 Flash in one chat: the idea, then 'make it alive', then 'the polish'. Three playable versions, each one a single HTML file.

3 prompts3 demos ๐Ÿ”’ members

AI tools: Step-5 vs GPT-6.1 โ€” playable games and real apps

Six original browser demos, exact prompts, measured run data and reproducible source code.

6 prompts6 demos ๐Ÿ”’ members

Claude Haiku 5.5 vs GPT-6 Luna, on the prompts Opus 5.5, Sol and Sonnet 5.5 got

Two small models with the same price per token, on launch day, given word for word the four prompts Opus 5.5 and GPT-6.1 Sol got on October 6 and the zombie-game prompt Sonnet 5.5 got on September 29.

5 prompts19 demos ๐Ÿ”’ members

VOXARA: one multiplayer voxel-sandbox prompt, five models

The same prompt, word for word, to Claude Opus 5.5, GPT-6.1 Sol, GPT-6 Luna, Claude Haiku 5.5 and Step-5 Preview: a peaceful multiplayer voxel world with biomes, a car, a boat, a plane, a build mode with prefabs and peer-to-peer multiplayer with a room code.

1 prompts5 demos ๐Ÿ”’ members

Mistral Large 4 vs Claude Opus 5.5 vs GPT-6.1 Sol: same four prompts

Mistral's one-trillion-parameter open model on launch week, given word for word the four prompts Opus 5.5 and GPT-6.1 Sol got on October 6: a voxel pagoda, a card machine, a working 3D kitchen and a racing game.

4 prompts12 demos ๐Ÿ”’ members

The $200 question: GPT-6.1 Sol vs Claude Opus 5.5, same four prompts

The models behind ChatGPT Pro and Claude Max on four prompts in the style of the leaked GPT-Next tests: a hover-racing game, a working 3D kitchen, a voxel pagoda and an isometric card machine.

4 prompts8 demos ๐Ÿ”’ members

Kolibri-1 vs Qwen, Nemotron and Mistral: same two prompts

Aleph Alpha's open German/English Kolibri-1 against four models from its own benchmark table: a neon hummingbird game and an interactive Mixture-of-Experts explainer, exactly the same prompt text for all.

2 prompts15 demos ๐Ÿ”’ members

The Fable 5.5 tests on Fable 5.1 and the free Ling 3.1 Flash

Four tests people ran on the leaked Claude Fable 5.5 (an art-history film, a forest jeep, a rope puzzle, a controller designer), given to Claude Fable 5.1 and to inclusionAI's free Ling 3.1 Flash.

4 prompts9 demos ๐Ÿ”’ members

Space Bunny vs MiniMax M3: island, shooter, desktop OS (MiniMax M3.1 Flash video)

Space Bunny Alpha (the free model the community links to MiniMax M3.1 Flash) and MiniMax M3 on the same prompts: a voxel island, a wave-survival shooter played by a bot, and a desktop OS run three times.

4 prompts6 demos ๐Ÿ”’ members

Cloudline: one prompt, five models (Gemini 4 Argon video)

The tram-game prompt YouWare gave Gemini 4 and GPT-6 Astra, run by us on Claude Sonnet 5.5, GPT-6.1 Sol and Gemini 3.8 Flash. Exact prompt, settings, costs and the three games' full code.

1 prompts3 demos ๐Ÿ”’ members

GPT-6.1 Sol: four demos + the same-prompt test vs Sonnet 5.5

A souls-like boss fight, a 3D keyboard assembly, a 360ยฐ watch viewer and a 3,600-line New York skyline, one prompt each, plus Sonnet 5.5's versions of the same two prompts.

4 prompts6 demos ๐Ÿ”’ members

Claude Sonnet 5.5: five demos

A zombie survival game, a GPU launch page, a 360ยฐ watch viewer, a 2,000+ line New York skyline and a voxel sandbox, each from one prompt, with the bots that played them on camera.

5 prompts5 demos ๐Ÿ”’ members

Space Bunny Alpha: six demos from a free mystery model

A Japanese garden, a shoot-'em-up, a headphones product page, a desktop OS, a GPU page and a flight globe, all written for $0 by the stealth model on OpenRouter.

6 prompts6 demos ๐Ÿ”’ members

Opening soon, by invitation.

All current and future projects, prompts, code and members' videos.

Join the waiting list