Ebitengine game

Deliver a game that is understandable, responsive and complete at the requested scope. One request can include several internal build/play/refine passes. A scaffold, successful build or attractive screenshot alone is not a finished game. For a focused fix, preserve the project's design and limit the workflow to the affected behavior.

The bundled starter targets Ebitengine v2.10.5 / Go 1.25+. This is a pinned baseline, not a claim about the latest release. Keep an existing project's versions unless an upgrade is needed and authorized. Check the pinned version's public API before using newer or experimental APIs. References contain patterns to adapt; only the bundled scripts and starter are executable deliverables. Their checks do not validate every reference snippet.

Delivery workflow

1. Inspect and decide. Read repository instructions, go.mod, the RunGame entrypoint and existing tests/assets. For new games, read Design and quality. Establish the player's goal, repeated decision, consequence, end/retry path, intended look, platform and supported input. Infer reasonable defaults; ask only for missing choices that materially change the product. Record a short brief and observable acceptance criteria in the project.
2. Build a representative playable slice. Make the central interaction and its feedback work with the intended controls and visual direction. Check the feel before expanding content. Keep scope small enough to finish; choose authored or procedural content based on the game, not the template. Calibrate content against actual player capabilities before multiplying levels. For an art set, establish shared perspective, scale, anchors and palette roles using Assets and animation.
3. Complete and refine. Implement the promised start → play → outcome → retry/continue loop (or the appropriate persistent/sandbox loop). Refine hierarchy, animation, sound, difficulty and transitions in context. Read only the technical references needed for the selected systems. Fix discovered defects within this delivery rather than asking the user to iterate for you.
4. Verify the product. Test rules and important transitions, exercise real input, inspect representative rendered states, and build/run requested targets. Include failure/retry and boundary cases. A scripted demo cannot establish control feel; a silent audio sink cannot establish sound quality. For automated multi-step interactions, pair observed images with actual state at the same tick; see Playtest diagnostics.
5. Hand off. Supply launch/build instructions, actual artifact paths and asset licences. State what was run and observed, what remains unverified, and any material limitations. Publishing, purchases and external uploads are separate actions within the user's requested scope.

Starting a project

Resolve commands relative to this skill's directory, not the game's working directory. Set skill_dir to the actual installed location, then for example:

python3 "$skill_dir/scripts/new_game.py" --out mygame \
  --module example.com/mygame --title "My Game"
cd mygame
go mod tidy
go test ./world/...

The starter demonstrates scenes, action input, text, sound, a camera, pause and scripted capture using a small platformer. It is an integration example, not a genre or architecture requirement. Replace its rules, controls, level, presentation and scene flow to fit the brief; a puzzle or card game should not inherit gravity, jumping, camera scrolling or a platformer HUD. Existing projects rarely need a scaffold. --ci adds a workflow and tools; inspect its triggers, targets and release behavior before enabling it.

Loop and resource contracts

Method	Responsibility	Important distinction
Update() error	Sample actions, advance simulation, apply scene changes and trigger event audio	Default target is 60 fixed ticks/s. Catch-up updates, focus pauses and slow machines affect actual scheduling.
Draw(screen *ebiten.Image)	Render the state produced by updates	Calls are not one-to-one with ticks. Do not advance gameplay, consume input edges or start sounds here. Renderer-owned caches are separate from simulation state.
Layout(outsideW, outsideH int) (int, int)	Choose logical dimensions for rendering and input	Fixed and responsive designs are both valid. Account for small windows, aspect ratio and readability.

• Express simulation durations in ticks or use an explicit fixed dt. At 60 TPS, 30 ticks is 0.5 seconds of simulation time. ActualTPS is a diagnostic, not a delta. SetTPS changes the whole update clock; it is not a universal slow-motion control. Avoid SyncWithFPS for rules that require a fixed step.
• Return ebiten.Termination from Update for a clean stop. Other errors return from RunGame; propagate failures rather than hiding them.
• Load/decode images and font sources once; reuse offscreens and hot-path slices. Profile before making pooling or zero-allocation rules universal.
• Separate rules from device/GPU code where that makes them testable. A small game need not grow an ECS, scene stack or service layer without need.

Essential Ebitengine choices

• Input: Map devices to actions in Update. Separate held/pressed/released; combine bindings before computing an action edge if that is the intended behavior. Standard gamepad mappings need an availability check and stick dead zones. Verify the advertised devices through menus and retry too. Input.
• Rendering: GeoM operations apply in call order. Pixel art needs deliberate sampling/scaling; smooth art needs suitable filtering and antialiased geometry. Preserve layer order while batching; sharing a sheet does not guarantee one command. Custom final-screen transforms need matching pointer conversion. Rendering.
• Text/UI: Preserve suitable existing bitmap/atlas text rendering. For new outline-font code, use text/v2 and parse font sources once. Include fonts covering the shipped scripts and their licences. Measure actual strings and verify focus, long text and target-size readability. Text/UI.
• Audio: One audio.Context per process. Match sample rate and PCM format: DecodeF32 does not resample; DecodeWithSampleRate returns 16-bit PCM. Overlapping effects need independent players or a voice pool; bound concurrency for dense events. Resume browser audio after a gesture. Audio.
• Systems: Tweens, particles, shake, physics and map loaders are optional tools supporting the central interaction. Gameplay systems.
• State: Queue external callbacks for Update, make reset/retry clear, separate gameplay and cosmetic randomness, test saves if persistence is promised. Architecture/testing.

Verification

Run commands from the game directory; retain go.mod and go.sum. Use repository build flags; -buildvcs=false below avoids depending on surrounding Git metadata.

go vet ./...
go test ./world/...   # starter's GPU-independent rules
python3 "$skill_dir/scripts/check_ebitengine.py" .
go build -buildvcs=false -o game .

Run the checker after changing loop/input/audio code or reviewing old APIs. It is a read-only, text-based advisory scan: inspect warnings in context. --json returns the shared checker schema; --strict fails on warnings and --max-issues N limits the report without changing totals. Deprecated APIs are warnings, not proof of a broken build. A clean result proves neither API correctness nor quality. Run other relevant package tests too; on Linux, tests that initialize graphics need xvfb-run -a.

For the starter, capture with -demo -shot frame.png -shot-tick 240. On Linux, use a virtual display and, if needed, null audio without changing HOME:

audio_config=$(mktemp)
printf 'pcm.!default { type null }\n' > "$audio_config"
ALSA_CONFIG_PATH="$audio_config" xvfb-run -a ./game \
  -demo -shot frame.png -shot-tick 240
rm "$audio_config"

Open the PNG. Inspect more states as needed: first play, a busy/difficult moment, failure and outcome. Adapt scripted actions/capture points after changing the game; the template demo no longer proves anything about a different loop. For deeper input/replay tests, v2.10's opt-in VM may help (see Architecture/testing).

Build only promised targets, for example:

GOOS=windows GOARCH=amd64 go build -buildvcs=false -o game.exe .
GOOS=js GOARCH=wasm go build -buildvcs=false -o game.wasm .

Cross-compilation establishes buildability, not runtime support. Run browser builds over HTTP and inspect console/input/audio when they are a target. Native SDKs, store bindings and cgo dependencies can add requirements beyond Ebitengine's pure-Go desktop backend.

Compose within this catalog

This skill owns Ebitengine/Go APIs. For portable design concerns, pair it with the relevant installed discipline (for example game-ui-ux, game-feel, create-game-assets, input-systems, performance-optimization or save-systems), genre or shipping workflow. Bind the engine slots to this skill's references, not another engine's example code. Load only what the task needs; the complete folder also works on its own.

Read as needed

• Design and quality: brief, genre choices, complete loop, presentation and playtest acceptance; for new games/redesigns.
• Project and builds: desktop, WASM and mobile.
• Rendering, input, text/UI, audio.
• Assets and animation: when making or importing an art set; shared visual rules, scale, pivots and sprite-sheet QA.
• Playtest diagnostics: when scripting interactions or reproducing failures; controlled ticks, state/image evidence.
• Architecture/testing and gameplay systems: state, saves, animation, collision and optional rigid bodies. Tiled maps covers a constrained loader only when the game uses Tiled.
• Networking: only for requested online play.
• Publishing/platforms, monetization, CI/release: only for requested distribution/integration work.

If available, Ebitengine's repository skills writing-kage-shaders and run-ebitengine-app-headless provide version-specific advanced workflows. Other game-design/store skills are optional supplements; this skill remains usable without an external skill catalog.