I had the idea in 2013, in ActionScript for Flash: sliding tiles you push together to build the hole. I built it over Christmas 2025, scrapped that, started again in April, and shipped it — a mini-golf game with a level editor, a 108-hole campaign, multiplayer, and a 3D course I had to rebuild to fit on a phone.
I had this in ActionScript for Flash Player, in 2013. One mechanic: tiles you slide into place to build a course, and the course you build is the hole you play.
That's still the whole game. Every draggable tile and insert block in Putt Panic now comes from it.
Adobe killed Flash Player at the end of 2020, so I ended up building it on something else entirely — twice, as it turned out.
Slide a tile · finish the lane · play it
The git history starts on 29 April 2026, but the work doesn't. I went at this properly over Christmas 2025, then scrapped it and started again in April.
Both clips are the real game, captured frame by frame off the build.
June is when it stopped being a sandbox and turned into something I could ship.
Busiest day was June 30 — 50 commits, the day I tore the renderer out and rewrote it. That one's below.
In the order it actually happened, with what went wrong at each step.
First commit already had water, walls, sand, and tiles that slide — not a standing start, I'd built most of it twice by then. No menu, no title screen, no UI, just a playable board in a browser tab. Everything else got built around that physics loop.
phasercustom physicstile editorMay was the boring part: orientation lock, iOS scaling, a real HUD, icons, a splash screen, a first-run tutorial. The game is forced landscape, so every drag goes through a coordinate remap. Nobody notices that unless it's broken.
capacitorforce-landscapeonboardingThe level editor and multiplayer landed close together. Free-play doubles as the obstacle showcase, so I only maintain one of them. Multiplayer runs on Firebase, and shaped courses have to ship their whole board mask or the other phone rebuilds the wrong hole.
firebase roomsdraggable tilesboard masksI didn't draw the campaign. It's 108 generated courses baked to JSON — families that cycle, signature holes at 9 and 18, par medals, and a bench of rival scores to chase. The generator is unit-tested, and I re-baked all 108 enough times that "re-bake campaign" is its own genre in my git log.
procedural generationvitestmedals + rivalsThe courses render in 3D. On desktop, running Three.js in a second WebGL context next to Phaser was fine. On a 3GB iPhone it killed the app — white screen mid-render. It wasn't device RAM. It was WebKit's per-process GPU ceiling, which you can't buy your way past.
A smaller texture wasn't going to fix it, so I took Three.js out of the runtime. On June 30, in 50 commits, I replaced it with a Canvas2D shell that draws the board inside Phaser's one GL context off a baked projection. Three still bakes the look offline. The phone never pays for a second renderer.
three.js → canvas2dgpu budgetbaked projectionGetting the wooden edge rails right took 32 commits in one day: cream vs ivory, oak grain, square butt joints instead of a 45° mitre, flush corners, the skirt on diagonal edges, the shadow in the corner. It's one detail. It's also what stops the board looking like a prototype.
it's always the cornerswall labThe menu sky runs off the actual time of day — dawn through dusk on one deterministic clock, with its own light arc per campaign cup. The windmills run on that same clock, so they're timing gates now instead of decoration.
time-of-daydeterministic clockambient fxThe last stretch was batches: a rebuilt results card, per-hole leaderboards, a redesigned obstacle guide, Speed Golf, wind on free-play and campaign courses, and an IAP layer stubbed for RevenueCat. The Shop stays hidden until real purchases work.
testflightleaderboardsiap (staged)The ones where the symptom and the cause were nowhere near each other.
Courses white-screened on iPhone XR — no JS error, nothing in the console, because iOS was killing the whole WebContent process. idevicesyslog showed the OOM. The ceiling is WebKit's per-process GPU budget, and it varies by device — so I wrote a native DeviceMemory probe to read the actual RAM and tier the 3D budget per device, then cut peak allocation until the XR held.
Geometry masks silently stopped working on WebGL in Phaser 4. setMask just did nothing, so I rebuilt shaped ramps as a silhouette plus fill strips.
The course — rails included — rendered through Three.js in a second WebGL context beside Phaser's. 3GB iPhones couldn't hold both: the second context kept getting killed, taking the board with it. That's what forced the rewrite — draw everything in Phaser's one GL context and never pay for a second renderer.
Backing out of multiplayer would drop me straight into the Speed Golf hole browser. Returning to the menu passed Phaser undefined for the no-options case — and Phaser only replaces a scene's stored data when you hand it something truthy. So the previous launch's payload just stayed put: one trip back from a Speed Golf run stamped {openSpeedGolf} onto the menu, and every visit after that re-delivered it. Always pass a concrete object.
Tapping a tile made every other movable block disappear. The guide layer hid all non-selected blocks on select or hover, so a single touch emptied the board of the pieces you were supposed to be sliding. They stay drawn in an idle style now.
I kept 115 notes as I went, one fact each, so I didn't lose two days to the same wall twice. Most of this page came out of them.
Long before either of the builds above, there was this. March 2023: a stock photo of a golf course, a clip-art driver, and a ball the size of a grapefruit — served from 127.0.0.1:5173. It still runs on that same port today.
Putt Panic has a level editor, a 108-hole campaign, multiplayer, time-of-day skies, and a 3D course I rebuilt to fit inside a mobile GPU budget. I built all of it myself, start to TestFlight.