testlab

The impossible-prompts library. Every episode I throw a batch of one-shot prompts at the newest AI model — no follow-ups, no retries... or sometimes... But it's all transparent for you to watch — so everything lands here: the exact prompt, and the unedited code it produced. Copy any prompt, (Yoink that shit) run it on whatever model exists when you're reading this, and compare.

This library grows every episode. Prompts are numbered globally and never renumbered — #01 will still be #01 at prompt #500.

Episodes

ep1 — I Gave Fable 5 Six Impossible Prompts (One Shot Each)

Watch: https://youtu.be/hsLi_DgYb-k

Filmed before this repo existed — the take-home prompts are preserved here; the builds lived and died on camera.

#	Prompt	What it is
01	GTA 5 sim	top-down open-world game, one HTML file
02	Seiko store	premium watch reseller storefront
03	YouTube Factory	Factorio-style channel sim on real data
04	Drum kit	playable kit + beat recorder, Web Audio
05	Rally sim	mobile-first dirt-stage racer
06	Intelligence Engine	creator-business dashboard on real stats

ep2 — I Gave Fable 5 Nine Impossible Prompts (Even Harder)

Watch: (video coming — link lands here at publish)

Nine prompts fired simultaneously into nine parallel agents. Three of them (09, 12, 14) were built into my real production SaaS — those ship as patches of every line the model wrote, plus the new files; point the same prompt at your own codebase.

#	Prompt	What you get	Run it
07	CS:GO-style shooter	one HTML file	open index.html
08	The internet in 1999	one HTML file	open index.html
09	Feedback Lab	patch + tests + seed data	apply to your own app
10	Tinder for video ideas	web app + my real ideas deck	node server.js or open index.html
11	Cabin restoration game	one HTML file	open index.html
12	Testing Lab	patch + template + tests	apply after 09
13	VTuber studio	one HTML file + pixel avatar	open index.html, allow webcam
14	Writing Mode (Google Docs yoink)	patch + tests	apply after 12
15	Claude Code Mission Control	node app	node server.js

Ship or Skip ep1 — I Built 6 AI Tools for My Business

Watch: (video coming — link lands here at publish)

A different format from the impossible-prompts episodes: six real tools for my own business, each one brief-brainstormed with the agent first (Claude Opus 4.8), built live, then judged ship or skip. Two were built into a private production SaaS and ship as a patch / prompt-only — point the same prompt at your own codebase.

#	Prompt	What you get	Run it
16	AI voice sales agent	Flask app + ElevenLabs agent + conversion dashboard	pip install -r requirements.txt && python app.py
17	ScreenPost (web + Electron)	Electron app + full Vitest suite	npm install && npm run dev
18	ScreenPost.io landing page	Next.js 16 + Tailwind v4 site	npm install && npm run dev
19	Feedback Lab — picture mode	prompt only (private SaaS)	apply to your own app
20	AI Mate OS dashboard	Flask everything-dashboard	pip install -r requirements.txt && python app.py
21	Send-to-CreatorGrowth	patch + tests (private SaaS)	apply to your own app

fusion vs opus 4.8 — Same 6 Prompts, Two Models, Head to Head

Watch: (video coming — link lands here at publish)

Same six one-shot prompts fired at two models at once — openrouter/fusion (a multi-model composite: panel of models + web search, then synthesized) vs anthropic/claude-opus-4.8 (a single frontier model) — captured with cost, latency, and the unedited builds, then judged side by side. All self-contained, single-file builds.

#	Prompt	What it is
22	GTA v2	top-down open-world game with weapons, one HTML file
23	The internet in 1999	dial-up boot + a browsable '90s web, one HTML file
24	Pixel Synth	FL Studio-style 8-bit chiptune looper, Web Audio
25	Flight Sim	3D jet + air combat, one HTML file (CDN three.js ok)
26	Destruction Sandbox	physics mayhem playground (CDN Matter.js ok)
27	Productivity Hub	Kanban + calendar + to-do + Pomodoro, one HTML file

This episode ships both builds per prompt — opus/index.html and fusion/index.html in each folder — plus the A/B harness (ab_test.py + viewer.html) used to fire and compare them.

five-agents-one-afternoon — Five Agents, One Codebase, One Afternoon

Watch: (link at publish)

Five coding agents run in parallel on the same private SaaS — one git worktree per agent, one one-shot brief each — then the good branches get merged. The gift is SWARM_SETUP.md: the isolated-worktree-per-agent pattern, so you can run your own swarm on one repo. Each feature ships as a patch of only that agent's work (private SaaS — the prompt is the product, the patch is the proof). Point the same prompt at your own codebase.

#	Prompt	What you get	Run it
28	Creator Intelligence dashboard	patch (route + aggregator + dark SVG dashboard)	apply to your own app
29	Token budget + tiers + access gate	patch + full test suite (metering, 402 enforcement, allowlist gate)	apply to your own app
30	X Inspiration	clean single-feature patch + tests (twitterapi.io → idea cards, Pillow card image)	apply to your own app

opus 4.8 vs glm 5.2 — Same 6 Builds, Paid Frontier vs Free Open Model

Watch: (video coming — link lands here at publish)

The same six one-shot prompts fired at two models at once — anthropic/claude-opus-4.8 (the paid frontier, on a Max subscription) vs z-ai/glm-5.2 (a free open-weight model, run inside Claude Code by pointing it at z.ai). Same prompt, same harness, one shot each. Each folder ships both builds — opus/ and glm/. Real receipts: the whole GLM day cost $11.70; Opus would've been ~$55 at metered API rates (it ran free on Max). GLM matched or beat Opus on speed/lines on most — and hit its wall on the heaviest build (the video factory).

#	Prompt	What it is
31	3D Anatomy Explorer	rotatable muscular figure, hover-to-name, one HTML file (three.js)
32	FIFA Football	top-down football + real AI opponent, first to 3, one HTML file
33	X Video Factory	Tinder-swipe X feed → composited short-form preview (Node + ffmpeg)
34	Seiko 7S26 Movement	interactive 3D watch movement + assemble/explode slider (three.js)
35	AI Minecraft	voxel world — mine, chop a tree, craft a sword, kill a pig (three.js)
36	Text Write-On Animation	Remotion word-synced kinetic typography from a spoken line (bring your own audio)

Rules every episode runs under

• One shot each — no back-and-forth, no retries.
• All prompts in an episode fire at the same time into parallel agents.
• Libraries from a CDN allowed; bundled asset files not.
• Builds into private codebases ship as patches of every line the model wrote — the prompt is the product, the patch is the proof.
• Model is logged per episode (ep1–ep2: claude-fable-5; ship-or-skip-ep1 + five-agents-one-afternoon: claude-opus-4-8; fusion-vs-opus-4.8: openrouter/fusion vs anthropic/claude-opus-4.8; opus-vs-glm-5.2: anthropic/claude-opus-4-8 vs z-ai/glm-5.2).
• The Ship or Skip series is the exception to "one shot each": each build is brainstormed with the agent first, then built — the take-home prompt is the brainstorm-first brief.

MIT licensed. Yoink anything.