game-orchestra

Game Orchestra

The main orchestrator workflow for web game development. This is a reference document — the game-dev-agent reads and follows this decision tree. The agent performs the actions; this skill documents the workflow.

Entry Point

This skill is the entry point for all game development work. When invoked, the game-dev-agent reads this document and follows the decision tree below.

Pre-Flight

Before routing, the game-dev-agent MUST:

1. Read Claude MEM for user context and preferences
2. Read lessons-learned.md if it exists in the project
3. Read progress.md if it exists in the project
4. Assess project state (files, package.json, existing code)

Decision Tree

No design doc exists       → invoke game-designer skill
Design approved, no code   → invoke game-developer skill
Code exists, untested      → invoke game-qa skill
Tests pass, needs polish   → invoke game-artist skill
Code complete, pre-release → invoke game-reviewer agent
Any checkpoint reached     → invoke game-psychologist agent
Cycle complete / bug fixed → invoke game-retrospective agent
Session ending             → invoke game-retrospective agent
Psychologist FAIL          → invoke game-retrospective agent

Psychologist Checkpoints

The game-psychologist agent is invoked at these validation gates:

1. Post-design — validate game concept against learning science
2. Post-development — review interaction patterns and accessibility
3. Post-playtest — interpret QA results through psychological lens
4. Ad-hoc — when new content added, visuals changed, or user reports concern

Parallel Execution

Use dispatching-parallel-agents (existing CC skill) for independent tasks:

• game-designer and game-qa are always sequential (design before test)
• game-artist can run in parallel with game-qa ONLY when core mechanics are stable (artist changes should be CSS/style only to minimize rework)
• Psychologist checkpoint is sequential (must complete before next phase)
• Retrospective can run in background after cycle completion

Mandatory Independence Check

Before ANY parallel dispatch, the dispatching agent MUST output:

=== INDEPENDENCE CHECK ===
PARALLEL AGENTS: [agent-name-1], [agent-name-2]
AGENT-1 TOUCHES: [list of files/systems agent-1 will modify]
AGENT-2 TOUCHES: [list of files/systems agent-2 will modify]
FILE OVERLAP: [none | list of conflicts]
DATA DEPENDENCY: [none | "agent-2 needs agent-1 output" — describe]
DECISION: [PROCEED | SEQUENTIAL — reason]
=== CHECK COMPLETE ===

• If FILE OVERLAP is not none, dispatch MUST be sequential
• If DATA DEPENDENCY is not none, dispatch MUST be sequential (unless agents can work from shared context — state why)
• Use actual agent names, not placeholders

Progress Tracking

Only the game-dev-agent writes to progress.md. Parallel agents report results back to the game-dev-agent for recording.

See templates/ for progress.md format.

Information Capture & Validation

All dispatched agents MUST follow:

1. Information Capture Protocol (see references/information-capture.md)
2. 6x Validation Test (see references/validation-protocol.md)