Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Build a 2D platformer: run/jump control with coyote time, jump buffering, and variable jump height, plus tiled levels and hazards. Use for a platformer or Mario/Celeste-like, or tuning jump feel.
.claude/skills/gamedev-skills-platformer/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-12 | ✗→✓ | ▲ Improved | 117% | 0% |
| case-15 | ✗→✓ | ▲ Improved | 28% | 0% |
| case-19 | ✗→✓ | ▲ Improved | 33% | 0% |
| case-01 | ✓→✓ | = Same ✓ | 52% | 0% |
| case-02 | ✓→✓ | = Same ✓ | 74% | 0% |
A playbook for 2D platformers — the run/jump controller "feel", level structure, hazards, and goals. This is a compositional skill: it wires an engine movement skill, a tilemap skill, and design skills into a working game. It does not re-teach physics or tilemaps; it tells you what to build and how to make jumping feel good.
"Celeste-like", or any game whose core verb is jump between surfaces.
(coyote time, jump buffering, variable height, corner correction).
When not to use: top-down movement with no gravity → use the engine movement skill directly. 3D first-person traversal → fps-shooter. Grid/turn movement → roguelike. For the raw kinematic body API, use godot-2d-movement (or your engine's controller skill).
Observe a gap/hazard → commit to a jump or move → land safely (or die) → reach the next checkpoint/goal. A platformer lives or dies on the moment-to-moment feel of that single jump, repeated thousands of times. Tighten the controller first; everything else is content.
Tune these by outcome (height in tiles, time to apex in seconds), not by raw numbers.
| Knob | Effect | Sane starting point | |------|--------|---------------------| | Max jump height | reach | 3–4 tiles | | Time to apex | "weight"/snappiness | 0.30–0.40 s | | Fall gravity multiplier | snappy, non-floaty fall | 1.5–2.0× rise gravity | | Coyote time | jump just after leaving a ledge | 0.08–0.12 s (~5–7 frames @60) | | Jump buffer | press just before landing still jumps | 0.10–0.15 s | | Variable jump cut | tap = short hop, hold = full | cut upward velocity ×0.4–0.5 on release | | Apex hang | brief float at the top for air control | reduce gravity ×0.5 near |vy|<threshold | | Ground accel / friction | responsiveness vs. ice | reach top speed in 0.05–0.1 s | | Corner correction | nudge past a ledge clipped by 1–2 px | nudge up to ~4 px sideways |
Derive gravity and jump velocity from the feel values rather than guessing — see Pattern 1.
python# Pseudocode. Pick the FEEL you want, then derive the physics. y-axis points DOWN. # From kinematics: h = (g * t^2) / 2 and v0 = g * t. JUMP_HEIGHT = 3.5 * TILE # how high, in world units TIME_TO_APEX = 0.35 # seconds to reach the top gravity = (2 * JUMP_HEIGHT) / (TIME_TO_APEX ** 2) # rising gravity jump_velocity = -(2 * JUMP_HEIGHT) / TIME_TO_APEX # negative = upward fall_gravity = gravity * 1.8 # heavier on the way down → less floaty
python# Pseudocode in the per-frame update. dt = seconds since last frame. # Timers count DOWN; refresh coyote while grounded, buffer on a fresh press. if on_floor: coyote_timer = COYOTE_TIME # 0.1 if jump_pressed_this_frame: buffer_timer = JUMP_BUFFER # 0.12 coyote_timer -= dt buffer_timer -= dt # A jump is allowed if we pressed recently AND were grounded recently. if buffer_timer > 0 and coyote_timer > 0: velocity.y = jump_velocity buffer_timer = 0 coyote_timer = 0 # consume both so we can't double-jump # Variable height: releasing jump early while still rising cuts the arc short. if jump_released_this_frame and velocity.y < 0: velocity.y *= 0.45 # Asymmetric gravity: snappier fall than rise. g = fall_gravity if velocity.y > 0 else gravity velocity.y += g * dt
Solid from above, pass-through from below. Most engines expose a "one-way collision" flag on the tile/collider; enable it and let the player drop through by disabling that collision for a few frames when the player holds Down + Jump. Do not re-implement collision math.
dt → speed changes with frame rate. Every velocityintegration and timer must use dt. (See physics-tuning.)
per-tile colliders, and add corner correction.
timestep for fast bodies (see physics-tuning).
godot-2d-movement (Godot CharacterBody2D); for other engines usethe engine core + physics skill (unity-physics, phaser-arcade-physics, pygame-core).
godot-tilemap / unity-tilemap-2d for geometry; level-design for layout,pacing, and teaching order.
physics-tuning for timestep, CCD, and stability.input-systems for buffering, rebinding, and gamepad support.audio-design for SFX/music; the engine animation skill for squash/stretch.prototype-fast to greybox the controller before building content.platforms, and camera follow, read references/feel-tuning.md.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | pass→pass | 11,335 | 7,733 | -32% | 1 | 1 | 0% | 2,048 | 3,113 | +52% | 0 | 0 | — |
case-02 | pass→pass | 9,492 | 6,979 | -26% | 1 | 1 | 0% | 1,933 | 3,355 | +74% | 0 | 0 | — |
case-03 | pass→pass | 10,615 | 3,219 | -70% | 1 | 1 | 0% | 1,937 | 2,345 | +21% | 0 | 0 | — |
case-04 | pass→pass | 8,995 | 4,299 | -52% | 1 | 1 | 0% | 1,602 | 2,530 | +58% | 0 | 0 | — |
case-05 | pass→pass | 8,776 | 5,206 | -41% | 1 | 1 | 0% | 1,524 | 2,781 | +82% | 0 | 0 | — |
case-06 | pass→pass | 9,545 | 6,130 | -36% | 1 | 1 | 0% | 1,717 | 2,846 | +66% | 0 | 0 | — |
case-07 | pass→pass | 6,987 | 4,169 | -40% | 1 | 1 | 0% | 1,097 | 2,538 | +131% | 0 | 0 | — |
case-08 | pass→pass | 10,533 | 4,973 | -53% | 1 | 1 | 0% | 1,767 | 2,602 | +47% | 0 | 0 | — |
case-09 | pass→pass | 9,772 | 5,511 | -44% | 1 | 1 | 0% | 1,579 | 2,644 | +67% | 0 | 0 | — |
case-10 | pass→pass | 9,590 | 6,396 | -33% | 1 | 1 | 0% | 1,367 | 2,757 | +102% | 0 | 0 | — |
case-11 | pass→pass | 8,958 | 5,359 | -40% | 1 | 1 | 0% | 1,671 | 2,578 | +54% | 0 | 0 | — |
case-12 | fail→pass | 6,839 | 4,342 | -37% | 1 | 1 | 0% | 1,166 | 2,532 | +117% | 0 | 0 | — |
case-13 | pass→pass | 11,477 | 2,682 | -77% | 1 | 1 | 0% | 1,917 | 2,243 | +17% | 0 | 0 | — |
case-14 | pass→pass | 12,218 | 11,396 | -7% | 1 | 1 | 0% | 2,148 | 3,936 | +83% | 0 | 0 | — |
case-15 | fail→pass | 10,141 | 3,360 | -67% | 1 | 1 | 0% | 1,850 | 2,372 | +28% | 0 | 0 | — |
case-16 | pass→pass | 11,401 | 13,065 | +15% | 1 | 1 | 0% | 1,793 | 3,695 | +106% | 0 | 0 | — |
case-17 | pass→pass | 6,874 | 4,330 | -37% | 1 | 1 | 0% | 1,313 | 2,576 | +96% | 0 | 0 | — |
case-18 | pass→pass | 6,124 | 2,207 | -64% | 1 | 1 | 0% | 1,144 | 2,214 | +94% | 0 | 0 | — |
case-19 | fail→pass | 15,361 | 8,198 | -47% | 1 | 1 | 0% | 2,465 | 3,286 | +33% | 0 | 0 | — |
case-20 | pass→pass | 10,171 | 8,527 | -16% | 1 | 1 | 0% | 1,678 | 3,160 | +88% | 0 | 0 | — |
case-21 | pass→pass | 12,609 | 6,874 | -45% | 1 | 1 | 0% | 1,959 | 2,778 | +42% | 0 | 0 | — |
case-22 | pass→pass | 16,691 | 11,154 | -33% | 1 | 1 | 0% | 2,657 | 3,475 | +31% | 0 | 0 | — |
case-23 | pass→pass | 16,006 | 4,570 | -71% | 1 | 1 | 0% | 2,346 | 2,608 | +11% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 23 cases were attempted. The headline lift of +13 percentage points is the difference between those two pass rates over the 23 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.