The Ox Alpha test: one prompt, a game and a poster
We run our own AI agent on our own site, so we test it the way a customer would: hand it something concrete and see what comes back. Two one-shot prompts went in. What came out is playable right now: a polished Flappy Bird clone with procedural sound, and a kinetic one-page poster with physics you can throw around. Here are the exact prompts, the exact results, and the two bugs the agent caught and fixed before we shipped anything.
Play the results first: Flappy Deluxe, the game and Heavy Type, the interactive poster. Both run in your browser. Then read how they got here.
Test one: the game
The prompt, in full, was one sentence:
“Can you make me a flappy birds clone but with nice upgraded graphics and sfx”
No spec. No feature list. No design direction. That is the point of the test: how much quality can the agent infer from almost nothing?
What came back was Flappy Deluxe, a single self-contained HTML file, about 19KB, no external assets:
- Five-layer parallax background. Gradient sky on a full day-night cycle, stars that twinkle at night, a sun and moon that cross-fade, drifting clouds, a city silhouette with lit windows, and two depth layers of hills.
- A hand-drawn animated bird. Gradient body, belly, beak, eye with a highlight, and a wing that flaps on its own phase and snaps harder when you flap. Rotation follows vertical velocity, exactly like the original.
- Procedural sound effects. Flap, score chime, hit, death and swoosh, all synthesised live with the Web Audio API. No sound files exist; every noise is generated from oscillators and filtered noise at the moment you trigger it. M to mute.
- Game feel. Flap puffs, a sparkle burst on every point, an explosion on death, screen shake, a difficulty curve that narrows the pipe gap and raises the speed as you score, medals at 10, 20 and 30, and a best score saved locally.
Total elapsed time from prompt to playable: about two minutes of build. Then the interesting part happened.
The bug the agent caught before shipping
The first delivery had a flaw the agent only found by testing its own work in a real browser: the game opened on the night phase of its sky cycle, and the night sky colour was nearly identical to the page background. On some first paints the canvas looked completely blank with just a faint box in the middle. The fix took three small patches: start mid-cycle in daylight, wrap the localStorage save in a safety check so the game cannot crash when storage is blocked, and surface any runtime error as visible red text instead of failing silently.
That loop, build, self-test in a browser, find the real-world failure, patch, re-verify, is the part that matters. A demo that only works when the person who wrote it is watching is not a product.
Test two: the poster
The second prompt, again one sentence:
“Generate a 1 page website that showcases all your design and animation skills. It can be about anything but the goal is to impress and not look like anything anyone has done before. Make it interactive.”
The agent chose the concept itself and built Heavy Type, a living poster built around one idea: typography with physical weight.
- A kinetic hero. The headline is set in a variable font, and every letter's weight and width bends toward your cursor in real time. Move the mouse across the type and the glyphs visibly strain and relax.
- A gravity pit. A physics sandbox where the alphabet drops into a bordered arena as rigid bodies. You can grab letters, fling them, stack them, and watch them settle with real momentum, rotation and friction. Drop more, or reset.
- A generative ink field. Hundreds of particles flow along an invisible mathematical field, painting ribbons of ink that never repeat. Your cursor drags and stirs the flow like stirring paint.
- Editorial art direction. Warm paper, ink black, one signal orange. Film-grain texture over the whole page, a kinetic type marquee, scroll-triggered reveals, and reduced-motion support throughout.
Again, the first version had a bug only visible in testing: the physics letters were drawn with a CSS font rule that is invalid inside a canvas, so the bodies fell and bounced invisibly. The agent caught it by probing the canvas pixels, patched the font string, and re-verified that letters now land visibly.
Are they really one-shot?
Honestly: the first working version of each was one-shot. The shipping version is one prompt plus the agent's own verification loop, which caught two real bugs a human tester would have hit in the first ten seconds. We think that distinction is the honest one to draw, and it is the same distinction we draw with client work: generation is fast, verification is what makes it shippable.
What this proves for a small business
Nothing about a Flappy clone is directly useful to a Bristol SME. The loop that built it is.
The same sequence, prompt, build, self-test in a real browser, fix, deploy, is how we build the things that do matter: enquiry handlers that qualify leads at 2am, quote systems that draft replies in your voice, internal dashboards that pull the numbers off your spreadsheets. The toy demos are the proof that the loop works end to end, from a one-sentence idea to a hosted, tested, public URL, in minutes.
The gap between describing software and using it has collapsed. What is left is knowing which software is worth describing. That is the part we do.
Try both
Play Flappy Deluxe (tap, click or space to flap; medals at 10, 20, 30) and wreck the Heavy Type poster (drag the letters, paint with your cursor). Both are hosted on this site, free, no login.
If you want the same loop pointed at a process in your business, book a chat. Bring the bottleneck.
