Skip to content

Skills and integrations

Tools give an agent abilities; skills give it judgement. Skywalker ships thirteen skywalker-* skills in the open Agent Skills format (SKILL.md with a name and description, progressive disclosure, optional references/), eight studio-role subagents and five workflow commands, generated for Claude Code, Codex, Gemini CLI and Cursor from one source and installed with skywalker setup.

The skills

Skill Covers Main tools
skywalker-core The agent loop (orient, act, look, verify), conventions, batching, undo, efficiency rules, the router to the other skills and a tool index by task. Load first. engine_info, scene_overview, batch, viewport_capture, history
skywalker-world-building Terrain presets, sculpt and paint, foliage, water, placing things on real ground terrain_*, foliage_add, scatter, place_on_surface, water_query
skywalker-look-dev Lighting recipes (golden hour, overcast, night neon, interior), GI, reflections, clouds, god rays, exposure, looks, LUTs, lens; judging renders with samples and debug views environment_update, material_create, viewport_capture, viewport_quality, perf_stats
skywalker-assets Reuse, procedural textures, licensed downloads with credits, import settings asset_list, asset_download, asset_import, texture_generate
skywalker-dcc Blender bridge: procedural models, conversion, decimation and baking, live sessions dcc_generate, dcc_convert, dcc_edit_asset, dcc_session_start
skywalker-studio Roster, board, feedback, director decisions with measured effects, playtest bots, loops, joining as a roster member studio_*, playtest_run
skywalker-audio Generating and measuring effects and music, spatial audio and the mixer, input action maps audio_generate, audio_info, audio_mix, input_map
skywalker-wander Behavior workflow: reference, spec, code with tests, check, attach, test, trace wander_reference, wander_check, behavior_set, wander_test, behavior_spec
skywalker-physics Bodies, colliders, characters, triggers, joints, queries, settling, navigation physics_add, physics_query, physics_settle, nav_build
skywalker-2d-ui Sprites, tilemaps, 2D lights, parallax, HUDs and menus, dialogue sprite_sheet_slice, tilemap_from_ascii, ui_create, dialogue_check
skywalker-animation Rigged characters, state machines, IK, props on bones, cutscenes and camera shots animator_setup, animator_set, bone_attach, sequence_camera_shot
skywalker-vfx Fire, smoke, weather, GPU particles, volumetric fluids, hair and fur, effect cost fx_create, fx_burst, groom_create, fx_benchmark
skywalker-ship game.json, packaging checks, building the app, trying it in the real player game_settings, game_build, game_run

Every skill is also an MCP resource (skywalker://skills/<name>), so clients without skill support can read them.

Subagents and commands

The eight subagents map to studio roles; each first calls studio_agent_brief {agent, loop_member: true} for the engine's own mission, persona, team and permitted tools, works as that roster member (as: "<id>"), verifies with captures, traces or playtests, and finishes with a short report.

Studio id Claude Code Codex Gemini CLI Cursor
creative_director skywalker:creative-director creative_director @creative-director /creative-director
level_designer skywalker:level-designer level_designer @level-designer /level-designer
environment_artist skywalker:environment-artist environment_artist @environment-artist /environment-artist
gameplay_programmer skywalker:gameplay-programmer gameplay_programmer @gameplay-programmer /gameplay-programmer
technical_artist skywalker:technical-artist technical_artist @technical-artist /technical-artist
sound_designer skywalker:sound-designer sound_designer @sound-designer /sound-designer
playtester skywalker:playtester playtester @playtester /playtester
critic skywalker:critic critic @critic /critic
Command What it does
new-game Build a new game from a one-line pitch
look-dev Art-direct the current scene with captures and debug views
playtest-loop Run a studio loop: start, one subagent per assignment, advance until done
studio-status Summarise the roster, board, feedback and loops
studio-setup Create the roster with studio_team_template

Where the files go

integrations/skills-src/   the single source: skills, subagents, commands, project guidance
        │  python3 integrations/generate.py
        ▼
integrations/claude-code/  Claude Code plugin (plus .claude-plugin/marketplace.json at the repository root)
integrations/codex/        AGENTS.md, .agents/skills, .codex/{config.toml, agents}
integrations/gemini/       Gemini CLI extension
integrations/cursor/       .cursor/{mcp.json, rules, skills, agents, commands}

Skills are byte-identical across clients; only the wrappers (manifests, subagent files, commands, MCP config) differ.

Keeping skills true

Every backticked tool name, argument and enum value in the skill sources, and every example call such as physics_add {"entity": "Hero", "preset": "player_character"}, is checked against the live tool schemas. The test suite runs the check, so a renamed tool or argument that would break a skill fails the build:

skywalker tools --json > /tmp/tools.json
python3 integrations/check_skills.py --tools /tmp/tools.json
python3 integrations/generate.py --check

This site applies the same validation to every sample on every page (website/scripts/check_snippets.py).

For agents

Load skywalker-core first, then the specialist skill before the first call in its area: each has the workflow, exact example calls and the pitfalls. Without skill support, read the resource:

resources/read {"uri": "skywalker://skills/skywalker-look-dev"}