Nine months ago I abandoned a Three.js game after three days of work. Last month, one session with Claude Fable 5 turned it into an actual game — and then the model went offline for 19 days before I could run a second session.
The idea started small: simple games I could someday play with my son. He's two. In September 2025 I gave myself three days to build an MVP of a 3D endless runner — a rabbit running through tunnels, collecting carrots — in pure Three.js, no game engine. The result was bad in ways I documented on the game's own dev blog at the time (the post-mortem is in Korean, but the before-screenshot speaks every language).
Three failures, specifically. There was no physics: the rabbit was glued to fixed points along a Catmull-Rom path — no gravity, no jump, no inertia. Rapier was declared in the code but never actually wired up. The visuals read as horror instead of cozy: an untextured Stanford Bunny floating in a black void. And there was no real gameplay — my own verdict was "a carrot-picking simulator." Every attempt to fix those things in raw Three.js hit a wall, and the repo went quiet from September 12 until June.
June 12: one session, 48 commits
Fable 5 launched on June 9, 2026. I had started a rebuild sprint with Claude Code on June 11 — new stack this time: Next.js 16, React 19, React Three Fiber v9, TypeScript, Zustand, TailwindCSS, and still deliberately no game engine. On June 12 I pointed Fable 5 at the project.
Forty-eight commits landed that day: physics integrated into the game loop, design docs plus implementation for the curved tunnel, a coin economy with an achievements registry, game UI (countdown, milestone banners, a slow-motion badge, a shop, daily records), trampolines and mid-air entities with collision handling, and tests to hold it all together.
The volume isn't the story, though. The story is which things worked on the first try. The physics feel — gravity, jump, and inertia while the rabbit runs along tunnel walls — came out right in one shot. So did procedural map design (maps generated from a daily seed, identical for every player) and the leaderboard on Upstash Redis. Those three systems were exactly what I had been stuck on for months in pure Three.js. I asked; they existed.
To be fair about attribution: this was a sprint, and Opus 4.8 and Sonnet did plenty of work around it. But the parts that had kept the project frozen for nine months came unstuck in that one Fable 5 session.
Then the model vanished — the same day
June 12 was also the day US export controls hit Anthropic's newest models, and Fable 5 went offline globally. It stayed offline for 19 days. I got exactly one session. I wrote up the shutdown and return separately: Claude Fable 5 is back online.
Work didn't stop, which is the useful part. Commit trailers from June 13–30 show Opus 4.8 — sometimes the 1M-context variant — and Sonnet 4.6 carrying the project; Sonnet 5 joined on June 30, its own launch day. The swap was painless because the project doesn't depend on any single model's habits. Conventions live in a project CLAUDE.md (file organization, naming rules), every change goes through a PR reviewed by Claude GitHub Actions — a flow that has been in place since day one in September 2025 — and jest plus husky gate what lands. Any model that respects those rails produces mergeable work.
When Fable 5 redeployed on July 1, it rejoined a moving project rather than rescuing a stalled one. The polish sprint since then has shipped three visual-polish passes, login streaks, daily quests, and a permanent upgrade shop — 24 commits on July 3 alone.
What didn't come free
Mobile optimization was the exception. It took several rounds of re-prompting — I sent it back again and again before phones ran the game acceptably. That fits a pattern I now plan around: agentic coding one-shots well-scoped, self-contained systems with clear inputs and outputs (physics, map generation, a leaderboard), and iterates like the rest of us on anything entangled with real devices and rendering budgets.
Before and after
| September 2025 MVP | Rebuild (June 2026 → now) | |
|---|---|---|
| Physics | None — rabbit pinned to points on a Catmull-Rom path; Rapier imported, never wired up | Custom gravity, jump, and inertia while running on tunnel walls |
| Visuals | Untextured Stanford Bunny in a black void | Cozy tunnel world; three polish passes so far |
| Gameplay | "A carrot-picking simulator" | Endless runner: 3 hearts, combo up to x5, rocks and roots, daily seeded maps |
| Stack | Pure Three.js | Next.js 16, React 19, React Three Fiber v9, TypeScript, Zustand, TailwindCSS |
| Backend | None | Upstash Redis leaderboard, daily records, coin economy, achievements |
Where the game is now
It's live at rabbit-hole-games.com. Moka the rabbit runs the tunnel walls, collects carrots, dodges rocks and roots. Everyone plays the same daily seeded map, so the leaderboard is a real comparison, not a lottery of layouts. The repo sits at 363 commits and roughly 37.7K lines of TypeScript. It runs free in the browser — play today's map and see where you land.
My son is still two, so the target audience hasn't reviewed it yet. But for the first time, the game looks like something I'd actually hand him.
What I'd tell you if your project is stuck
One session really can un-stick a nine-month-dead project — but the session wasn't magic on its own. I had already written a precise post-mortem of why the old version failed, and that document effectively became the spec. If your abandoned project has a written diagnosis, that's worth more than any prompt trick.
The rest is boring and repeatable: give the agent well-scoped subsystems, keep tests and CI in the loop so any model's output gets checked the same way, and don't couple your project to one model — mine survived losing its best one for 19 days without slowing down.
FAQ
Did Claude Fable 5 build the whole game?
No. The rebuild spans 363 commits across three weeks, and Fable 5 got exactly one day of it before export controls took it offline. Opus 4.8 and Sonnet carried the following 19 days. What made the one session notable is that it solved the specific systems — physics feel, procedural maps, leaderboard — that had blocked the project for nine months.
Do you need a game engine for a 3D browser game?
Not at this scope. The rebuild uses React Three Fiber v9 on React 19 and Next.js 16, with hand-written physics and Zustand for state. An engine buys you physics and an editor, but a web-native stack means deployment, auth, and the leaderboard are ordinary web problems — and an agent can work on all of it in one repo.
How do the daily seeded maps work?
The map generator is deterministic: each day's map is produced from a shared seed, so every player runs an identical layout. That makes daily leaderboard times directly comparable. This was one of the systems Fable 5 implemented in a single pass.