Glider through Time

The brief that started everything was disarmingly personal: "I am not a game developer. I do not speak the language. I want you to help me develop a 3D game that rivals professional ones." The inspiration was The Time Tunnel, a 1966 TV series where scientists fall through time into famous catastrophes and have to survive them.

The first attempt chased that literally — a Titanic narrative-survival game in a single HTML file. You'd materialise aboard the ship at 9 PM knowing the iceberg hits at 11:40, and your only weapon was foreknowledge: earn a stranger's trust with a small true prediction, spend it to pull them toward a lifeboat. It had a clever "credibility" mechanic, a retro-broadcast art style, a real countdown clock.

It was also, in the end, very disappointing. Too much game and not enough clarity — no hints, no obvious way to play, and low-poly geometry that couldn't carry the drama it promised. The verdict was one line: "very disappointing." We walked away.

The restart brief was vivid and specific: a silent hero glides across space-time on a hang-glider, seeing the great historical events, sites and personalities from above. It should be fun and realistic — explicitly "not blockish." That last word was a quiet warning we'd spend the whole project learning to respect.

Before a line of code, we ran a structured brainstorm — the AI asking one question at a time, the human steering. Three forks decided the shape of everything.

Offered a tidy six-scene grand tour, the human went bigger in the best way: "I want the entire history arc from primordial soup to the Second World War." The AI expanded it to a 54-scene itinerary across five acts — deep time, the rise of humanity, the classical world, the middle ages, the modern era — but flagged the cost honestly and proposed tiering: a dozen "hero" scenes with crowds and animation, the rest lighter atmospheric flythroughs.

Then, after an honest reckoning (next waypoint), the same ambition was disciplined down to a lean 10-scene first cut — one unforgettable beat per era, ending not at WWII but at the Wright brothers' first flight in 1903. The hang-glider hero meeting the birth of powered flight was a better ending than a blackout over a bombed city — and it sidestepped a real tonal problem we'll come to.

With a 16-phase implementation plan written, the human asked a pointed question: "I want Fable 5 to implement it. What elements would Fable 5 likely object to?" — inviting the model to critique its own work before committing. The honest objections mattered more than the plan did:

Every one of these came true. Naming them early is why the project bent gracefully instead of breaking at scene 40.

The lean cut came together in code: a Vite + Three.js project, a first-person cockpit rig, a real game HUD (compass, speed and altitude dials, minimap), procedural sky, terrain and water, a continuous flight path threading all ten eras, and — the validated bet — six free, rigged, animated CC0 dinosaurs sourced and dropped into the Cretaceous scene. It flew. Free-look worked. The captions faded in on cue.

And from the glider, the world below was tan pyramids made of cones, capsule-shaped crowds, cone-shaped trees on empty green hills. Technically a success; emotionally, a slideshow of primitives. Exactly the "blockish" the brief had warned against — and exactly the ceiling the pushback had predicted.

The references were stills from a hang-glider-over-fantasy-worlds clip — lush, cinematic, unmistakably photoreal. The instruction was blunt: "the graphics are disappointing… I want photorealistic visuals like those."

Reference frames above are stills from a video posted by @tyrtyre201 on X, supplied by the project owner as the target look. Shown here for comparison only; all rights remain with the original creator.

Here the AI had to push back on the customer, not the plan. The honest answer: no amount of code would get a real-time browser to that image — because that image was never rendered in real time. It was AI-generated video. Chasing it in Three.js meant polishing forever toward a ceiling that sat far below the target.

So the real question was never "write better graphics code." It was "which photoreal source do we use?" — and that reframing is what turned a disappointing game into a finished film.

Left to themselves, image models produce beautiful empty landscapes — a lone dinosaur on a vast plain — because "a T. rex in a field" is easier to draw than three hundred of them. The human named this failure mode before we hit it: "populate every scene… herds and herds of dinosaurs, a packed Colosseum. Otherwise it's just a video of greenery." So density had to be engineered into every stage, not hoped for.

The method, in four moves

1 · A population spec per era. Every prompt carried a counted, layered inventory — "hundreds of sauropods stretching to the horizon, dozens of triceratops in the foreground, a T. rex hunting at mid-distance, the sky full of pterosaurs" — because models respond to numbers and depth, not to the word "many."

2 · Keyframe first, cheaply. Each scene began as a single still image from Seedream V4 (three cents an image). Composition wrong? Retry the image, not the expensive video.

3 · A density gate. No still advanced to animation until it was genuinely full. This is where the human's eye ran the show — the first dinosaur frame came back with a suspiciously tidy single-file column, and the note was exact: "add some randomness and at least 5% more animals."

4 · Then, and only then, animate. The approved still was fed to Kling 2.5 Turbo Pro — an image-to-video model — with a written motion direction, producing a 5-second clip at $0.35 each.

First pass

Too orderly, too sparse — a single tidy column.

After the note

Chaotic herds on both banks, ~2× the density. Approved.

Wrestling the model into history

Photoreal models are confidently anachronistic. Giza came back with pyramid workers in modern clothes and — despite an explicit "no burning buildings" — a village on fire, because the model kept turning "cooking smoke" into flames. The fix was a genuine prompt-craft insight: mentioning "no flames" plants the very token you're trying to avoid. The winning move was to remove every fire and smoke word from the prompt entirely and let the village simply be quiet.

Fire the model invented

Modern dress, huts ablaze — from a prompt that asked for neither.

Tokens removed

Linen kilts, sledge crews, a clean village — after deleting the fire words.

Kitty Hawk had the opposite bug: the Wright Flyer kept fusing to the glider's own cables, dangling from us like a toy. An explicit "the aircraft flies entirely on its own, nothing attached to the glider" set it free.

All ten clips animated on the first try, so the entire retake budget went unspent. Assembly was free and local: ffmpeg laid era captions and instrument-style HUD pills over each clip, added a title card and a closing "The sky belongs to everyone," and generated a procedural wind soundtrack from filtered noise. But the first cut wasn't right, and each fix came from a specific human note:

Each was diagnosed openly (why it happened, options, cost), fixed for cents, and slotted straight back into the assembly. That loop — human notices, AI diagnoses and fixes, human judges — ran until there was nothing left to flag.

A non-coder who cannot "speak the language" shipped a photoreal historical film for the price of a coffee. That's only possible because the work split cleanly along a line: the human owned taste, direction, and truth; the AI owned research, building, and explanation. Neither could have done it alone — and crucially, the human never had to touch the code, the API, or the prompt syntax.

The through-line is that the AI was most useful not when it agreed, but when it pushed back with reasons — on the plan, on the graphics ceiling, on the platform — and when it turned each fork into a genuine choice with costs attached, rather than quietly picking one. The human's job was never to code. It was to decide, and to keep raising the bar until the thing was good.