Matt Sephton built BELTRUNNER for GMTK Jam 2026 as roughly 1,300 lines on Jinks, the small engine he is developing. It opens looking exactly like Asteroids, then drops the floor out: clear wave 1 and a black hole tears open where the next wave of rocks should be. What follows is a countdown race through numbered gates, in descending order, with time as the only resource.
The postmortem is unusually specific about which ideas earned their place:
- Masquerade first, then teach. Wave 1 is pure Asteroids — classic splits, classic thrust-and-drift, no lives at stake. It gets you flying on muscle memory before the race asks for anything harder.
- One mechanic per act. Sixteen waves as four acts of four. Each act introduces a single new gate behaviour — elliptical, then directional, then spinning — and spends three more waves letting it bed in. “A mechanic shown once and then immediately buried under the next one was never really taught.”
- Determinism as a design tool. Rock spawns, gate placement and power-up scatter all come from one seeded stream, so a single number describes the whole layout. That makes courses learnable, replays possible, and tuning a matter of adjusting rules rather than hoping.
- No lives. Time becomes the only resource, which deletes a pile of arcade bookkeeping and leaves one number that reads at a glance.
- Risk you can measure. FIELD is the live sum of invisible rock weights (large 1, medium 2, small 3), and gate payouts scale with it. Splitting rocks to clear a lane also raises the stakes.
- Secrets taught by ear. The rock field quietly highlights one rock at a time. Split them in order and sustained voices stack into a chord; split out of turn and the choir cuts out. A missing pip is easy to miss, a chord that stops is not.
The engine work is as concrete as the design work. Circle colliders shipped in the jam build and playtested badly — “the ship was a slender dart wearing a fat disc” — so they were replaced with concave polygon colliders and oriented ellipses that rotate with heading. Screen wrap became a real torus, because the fastest route to the next gate often runs through the screen edge and the player has to be able to trust it.
The takeaway is that arcade fidelity is a method, not nostalgia. Determinism, a single resource and stepped difficulty are doing the teaching here — no tutorial required.
Worth reading for what it documents alongside what worked: an early orbital prototype that was not fun, pre-maturely timed warp power-ups, a mean version of the freeze power-up, and the generalisation of every fix back into the engine.