Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use when implementing cameras — smooth follow, screen shake, camera zones, and transitions for 2D and 3D
.claude/skills/jame581-camera-system/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | 115% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 105% | 0% |
| case-21 | ✓→✗ | ▼ Worse | 106% | 0% |
| case-12 | ✓→✓ | = Same ✓ | 112% | 0% |
| case-02 | ✓→✓ | = Same ✓ | 155% | 0% |
All examples target Godot 4.3+ with no deprecated APIs. GDScript is shown first, then C#.
> Related skills: player-controller for first-person camera setup, state-machine for camera state transitions, godot-optimization for camera culling and performance, physics-system for physics interpolation and camera smoothing, 2d-essentials for canvas layers, parallax scrolling, and coordinate conversion, math-essentials for smoothstep and lerp-based interpolation, tween-animation for camera shake and cinematic transitions, phantom-camera for a Cinemachine-style camera addon.
| Property | Type | Description | |---|---|---| | position_smoothing_enabled | bool | Enables built-in position smoothing (lerp toward target) | | position_smoothing_speed | float | Speed of built-in smoothing (default 5.0) | | drag_horizontal_enabled | bool | Enables a horizontal drag zone; camera only moves when target exits zone | | drag_vertical_enabled | bool | Enables a vertical drag zone | | limit_left | int | Left pixel boundary — camera will not scroll past this | | limit_right | int | Right pixel boundary | | limit_top | int | Top pixel boundary | | limit_bottom | int | Bottom pixel boundary | | zoom | Vector2 | Zoom level; Vector2(2, 2) = 2× zoom in, Vector2(0.5, 0.5) = zoom out |
Set limits to match your tilemap or level bounds so the camera never shows outside the world. Limits are in world pixels, not tiles.
A manual follow camera gives more control than the built-in smoothing — you can add look-ahead, offset, and custom easing.
gdscriptextends Camera2D ## Target node to follow (assign in Inspector or via code) @export var target: Node2D ## How quickly the camera catches up to the target (higher = snappier) @export var follow_speed: float = 8.0 ## How far ahead the camera leads in the movement direction @export var look_ahead_distance: float = 80.0 ## How quickly the look-ahead offset responds to direction changes @export var look_ahead_speed: float = 4.0 var _look_ahead_offset: Vector2 = Vector2.ZERO var _previous_target_pos: Vector2 = Vector2.ZERO func _ready() -> void: # Disable built-in smoothing — we handle it manually position_smoothing_enabled = false if target: _previous_target_pos = target.global_position global_position = target.global_position func _process(delta: float) -> void: if not target: return # Compute movement direction from last frame var move_delta: Vector2 = target.global_position - _previous_target_pos _previous_target_pos = target.global_position # Smoothly steer look-ahead offset toward movement direction var desired_ahead: Vector2 = move_delta.normalized() * look_ahead_distance if move_delta.length() > 0.5 else Vector2.ZERO _look_ahead_offset = _look_ahead_offset.lerp(desired_ahead, look_ahead_speed * delta) # Lerp camera position toward target + look-ahead var desired_pos: Vector2 = target.global_position + _look_ahead_offset global_position = global_position.lerp(desired_pos, follow_speed * delta)
csharpusing Godot; public partial class SmoothFollowCamera : Camera2D { [Export] public Node2D Target { get; set; } [Export] public float FollowSpeed { get; set; } = 8.0f; [Export] public float LookAheadDistance { get; set; } = 80.0f; [Export] public float LookAheadSpeed { get; set; } = 4.0f; private Vector2 _lookAheadOffset = Vector2.Zero; private Vector2 _previousTargetPos = Vector2.Zero; public override void _Ready() { PositionSmoothingEnabled = false; if (Target != null) { _previousTargetPos = Target.GlobalPosition; GlobalPosition = Target.GlobalPosition; } } public override void _Process(double delta) { if (Target == null) return; float dt = (float)delta; Vector2 moveDelta = Target.GlobalPosition - _previousTargetPos; _previousTargetPos = Target.GlobalPosition; Vector2 desiredAhead = moveDelta.Length() > 0.5f ? moveDelta.Normalized() * LookAheadDistance : Vector2.Zero; _lookAheadOffset = _lookAheadOffset.Lerp(desiredAhead, LookAheadSpeed * dt); Vector2 desiredPos = Target.GlobalPosition + _lookAheadOffset; GlobalPosition = GlobalPosition.Lerp(desiredPos, FollowSpeed * dt); } }
A trauma-based system produces more natural-looking shake than a simple sine wave. High trauma = violent shake; trauma decays over time; offset scales with trauma^2 so small trauma values feel subtle.
gdscriptextends Camera2D ## Maximum pixel offset during maximum trauma @export var max_offset: Vector2 = Vector2(20.0, 15.0) ## Maximum rotation offset in degrees during maximum trauma @export var max_roll: float = 3.0 ## Rate at which trauma decays per second (0–1 range) @export var decay_rate: float = 1.5 var _trauma: float = 0.0 # 0.0 = no shake, 1.0 = maximum shake # Optional: use noise for smooth, organic shake var _noise: FastNoiseLite var _noise_time: float = 0.0 @export var use_noise: bool = true @export var noise_speed: float = 60.0 func _ready() -> void: _noise = FastNoiseLite.new() _noise.noise_type = FastNoiseLite.TYPE_SIMPLEX _noise.seed = randi() ## Call this from any node to trigger a shake (amount in 0–1 range; can stack) func add_trauma(amount: float) -> void: _trauma = minf(_trauma + amount, 1.0) func _process(delta: float) -> void: if _trauma <= 0.0: offset = Vector2.ZERO rotation = 0.0 return # Decay trauma over time _trauma = maxf(_trauma - decay_rate * delta, 0.0) _noise_time += delta * noise_speed var shake: float = _trauma * _trauma # squaring gives subtle feel at low trauma if use_noise: offset.x = max_offset.x * shake * _noise.get_noise_2d(_noise_time, 0.0) offset.y = max_offset.y * shake * _noise.get_noise_2d(0.0, _noise_time) rotation = deg_to_rad(max_roll) * shake * _noise.get_noise_2d(_noise_time, _noise_time) else: offset.x = max_offset.x * shake * randf_range(-1.0, 1.0) offset.y = max_offset.y * shake * randf_range(-1.0, 1.0) rotation = deg_to_rad(max_roll) * shake * randf_range(-1.0, 1.0)
Triggering shake from another node:
gdscript# Any node that can reach the camera func on_explosion() -> void: var cam := get_viewport().get_camera_2d() as ScreenShakeCamera if cam: cam.add_trauma(0.6)
csharpusing Godot; public partial class ScreenShakeCamera : Camera2D { [Export] public Vector2 MaxOffset { get; set; } = new Vector2(20f, 15f); [Export] public float MaxRoll { get; set; } = 3.0f; [Export] public float DecayRate { get; set; } = 1.5f; [Export] public bool UseNoise { get; set; } = true; [Export] public float NoiseSpeed { get; set; } = 60.0f; private float _trauma = 0f; private float _noiseTime = 0f; private FastNoiseLite _noise; public override void _Ready() { _noise = new FastNoiseLite(); _noise.NoiseType = FastNoiseLite.NoiseTypeEnum.Simplex; _noise.Seed = (int)GD.Randi(); } public void AddTrauma(float amount) { _trauma = Mathf.Min(_trauma + amount, 1.0f); } public override void _Process(double delta) { if (_trauma <= 0f) { Offset = Vector2.Zero; Rotation = 0f; return; } float dt = (float)delta; _trauma = Mathf.Max(_trauma - DecayRate * dt, 0f); _noiseTime += dt * NoiseSpeed; float shake = _trauma * _trauma; if (UseNoise) { Offset = new Vector2( MaxOffset.X * shake * _noise.GetNoise2D(_noiseTime, 0f), MaxOffset.Y * shake * _noise.GetNoise2D(0f, _noiseTime) ); Rotation = Mathf.DegToRad(MaxRoll) * shake * _noise.GetNoise2D(_noiseTime, _noiseTime); } else { Offset = new Vector2( MaxOffset.X * shake * (float)GD.RandRange(-1.0, 1.0), MaxOffset.Y * shake * (float)GD.RandRange(-1.0, 1.0) ); Rotation = Mathf.DegToRad(MaxRoll) * shake * (float)GD.RandRange(-1.0, 1.0); } } }
For room-based games (metroidvanias, top-down dungeons): an Area2D per room with a script that, on body_entered, tweens the active Camera2D's limit_left / limit_right / limit_top / limit_bottom to the room's bounds. Smooth transitions when the player crosses room boundaries.
> See references/camera-zones.md for the full GDScript + C# CameraZone implementation.
Three canonical 3D camera setups: third-person follow with SpringArm3D (handles wall collision), orbit camera with mouse-drag rotation, first-person with mouse-look-from-camera.
> See references/camera3d-patterns.md for full GDScript implementations of each pattern.
Async camera transitions via Tween + await ToSignal. The pattern: tween the next camera's position/zoom from the current camera's, then call make_current() on the next camera. Works for both 2D and 3D.
> See references/transitions.md for the full CameraTransitionManager implementation (2D and 3D).
Render multiple cameras to separate SubViewports, then arrange SubViewportContainers in a layout (HBoxContainer, VBoxContainer, or GridContainer). Each player's camera is set as current for its viewport.
> See references/split-screen.md for the SubViewport scene-tree setup and a 2-player split-screen example.
Camera2D limits match level/tilemap bounds so no out-of-world edges are visible_process (visual interpolation), not _physics_processoffset and rotation to zero when _trauma reaches 0.0add_trauma() clamps to 1.0; it does not exceed maximum shakeArea2D collision layers are set so only the player triggers themSpringArm3D collision mask includes all environment layers for wall avoidancemake_current()SubViewport sizes are updated on window resize (get_tree().root.size_changed signal)SubViewport has audio_listener_enable_2d or audio_listener_enable_3d set to true| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-12 | pass→pass | 13,627 | 9,723 | -29% | 1 | 1 | 0% | 2,561 | 5,423 | +112% | 0 | 0 | — |
case-01 | fail→pass | 14,805 | 13,785 | -7% | 1 | 1 | 0% | 2,883 | 6,201 | +115% | 0 | 0 | — |
case-02 | pass→pass | 9,532 | 8,225 | -14% | 1 | 1 | 0% | 1,910 | 4,871 | +155% | 0 | 0 | — |
case-03 | pass→pass | 17,731 | 7,546 | -57% | 1 | 1 | 0% | 3,033 | 4,476 | +48% | 0 | 0 | — |
case-04 | pass→pass | 14,345 | 11,100 | -23% | 1 | 1 | 0% | 2,438 | 5,163 | +112% | 0 | 0 | — |
case-05 | pass→pass | 13,105 | 6,343 | -52% | 1 | 1 | 0% | 2,385 | 4,412 | +85% | 0 | 0 | — |
case-06 | pass→pass | 4,595 | 4,814 | +5% | 1 | 1 | 0% | 823 | 4,113 | +400% | 0 | 0 | — |
case-07 | fail→pass | 12,628 | 7,793 | -38% | 1 | 1 | 0% | 2,232 | 4,573 | +105% | 0 | 0 | — |
case-08 | pass→pass | 14,478 | 16,870 | +17% | 1 | 1 | 0% | 2,468 | 6,207 | +151% | 0 | 0 | — |
case-09 | pass→pass | 14,751 | 11,293 | -23% | 1 | 1 | 0% | 2,445 | 5,161 | +111% | 0 | 0 | — |
case-10 | fail→fail | 12,732 | 12,396 | -3% | 1 | 1 | 0% | 2,159 | 5,402 | +150% | 0 | 0 | — |
case-11 | pass→pass | 10,028 | 9,535 | -5% | 1 | 1 | 0% | 1,594 | 4,798 | +201% | 0 | 0 | — |
case-13 | pass→pass | 17,033 | 9,375 | -45% | 1 | 1 | 0% | 3,086 | 4,912 | +59% | 0 | 0 | — |
case-14 | pass→pass | 16,782 | 12,796 | -24% | 1 | 1 | 0% | 2,656 | 5,394 | +103% | 0 | 0 | — |
case-15 | pass→pass | 4,562 | 2,772 | -39% | 1 | 1 | 0% | 701 | 3,643 | +420% | 0 | 0 | — |
case-16 | pass→pass | 15,254 | 14,177 | -7% | 1 | 1 | 0% | 2,540 | 5,753 | +126% | 0 | 0 | — |
case-17 | pass→pass | 13,949 | 11,307 | -19% | 1 | 1 | 0% | 2,336 | 5,104 | +118% | 0 | 0 | — |
case-18 | pass→pass | 6,575 | 5,107 | -22% | 1 | 1 | 0% | 1,086 | 4,108 | +278% | 0 | 0 | — |
case-19 | pass→pass | 12,793 | 9,049 | -29% | 1 | 1 | 0% | 2,013 | 4,651 | +131% | 0 | 0 | — |
case-20 | pass→pass | 12,141 | 10,657 | -12% | 1 | 1 | 0% | 2,417 | 5,185 | +115% | 0 | 0 | — |
case-21 | pass→fail | 16,819 | 15,409 | -8% | 1 | 1 | 0% | 2,977 | 6,128 | +106% | 0 | 0 | — |
case-22 | pass→pass | 12,523 | 13,612 | +9% | 1 | 1 | 0% | 2,182 | 5,699 | +161% | 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. 22 cases were attempted. The headline lift of +5 percentage points is the difference between those two pass rates over the 22 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.
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.