How It Works¶
GodotMaker is an optimized agent workflow for building Godot games from an idea. The CLI helps turn the idea into a GDD, drives specialized roles, then loops through verification, gameplay evaluation, screenshot review, and fixes until the current design scope is complete.
flowchart TD
A[Game idea] --> B[godotmaker-cli]
B --> C[GDD.md and current tag plan]
C --> D[Generate or inspect assets]
D --> E[Build with worker agents]
E --> F[Unit tests and verifier checks]
F --> G[Run game and E2E tests]
G --> H[Screenshot and visual QA]
H --> I{Matches GDD?}
I -- yes --> J[Finalize tag]
I -- no --> K[Fix gap]
K --> F
The old manual mode still exists: advanced users can run /gm-* commands directly. The product direction is CLI-driven no-human-in-the-loop execution.
The Four Phases¶
1. Plan¶
The workflow starts from your game idea. GodotMaker helps capture that idea as GDD.md, then turns the design into:
PLAN.md- tasks for the current tagSTRUCTURE.md- expected components, systems, and project structureSCENES.md- scene responsibilities and acceptance criteriaASSETS.md- required generated or provided assets
The GDD remains the design contract after idea capture. If you want a different game, refine the idea or edit the GDD and run the workflow again.
2. Build¶
Implementation is delegated to worker agents. Workers write game code and tests in scoped tasks rather than one large generation pass.
The build loop is deliberately strict:
- workers implement tasks from
PLAN.md - unit tests are written with the code
- a verifier runs headless Godot checks and gdUnit4
- reviewers check Godot-specific pitfalls such as physics, UI layout, animation, navigation, audio, shaders, particles, and tilemaps
This is why runs take hours instead of minutes: GodotMaker spends the time on verification and correction.
3. Evaluate¶
Verification proves that the project compiles and tests pass. Evaluation asks a different question: does the game actually match the design captured in the GDD?
The evaluator:
- launches the game
- writes or runs end-to-end tests
- simulates player-like actions
- captures screenshots
- checks visual and UI issues against the design
- writes structured findings
This catches problems that compile-time checks miss: overlapping UI, missing game-over flow, unreadable prompts, broken progression, or visuals that do not match the scene requirements.
4. Fix and Finalize¶
When evaluation finds a gap, /gm-fixgap creates a fix plan and dispatches workers to close it. The workflow then loops back through verification and evaluation.
When the current design scope passes, GodotMaker finalizes the tag:
- archives planning docs under
docs/tags/<Tag>/ - writes
.godotmaker/final_report.json - records a local git tag
- resets per-tag runtime state
You can then refine the idea or edit the GDD and start the next tag.
What Makes It Different¶
No-Human-In-The-Loop by Default¶
GodotMaker is designed so the CLI can keep working after the idea has been clarified into a design. The agent asks for human input when the design itself is unclear, but implementation, tests, evaluation, screenshots, and fixes are intended to run automatically.
TDD and E2E Are Part of the Workflow¶
Tests are not optional cleanup. Unit tests and E2E tests are part of the generation loop and become feedback for later agents.
Visual Feedback Is a First-Class Gate¶
Godot games can pass tests and still look wrong. GodotMaker captures screenshots and uses visual QA to turn UI and scene problems into concrete fix tasks.
Real Godot Output¶
The output is a normal Godot project. You are not locked into a hosted editor or a proprietary runtime.
For command-level details, see The 9 roles.