Make a Game from Zero with AI Agents: Six Stages and the Skills for Each
A route from one sentence to a Steam release for people who do not write code: what to ask the agent at each stage, what to check yourself, and which tested skills to load.
You can now take a game from an idea to a store page without writing code yourself, if the agent does the typing and you do the deciding. A solo developer with no coding background recently wrote up how they shipped a free early-access Steam game this way (the write-up, in Chinese). This page is our own route through the same stages, with the GameSkill skills that cover each one. Every linked skill page says what we tested and what we did not.
Two habits help at every stage:
- Keep one
AGENTS.mdat the project root with the rules every agent must follow (engine version, folder layout, "never edit generated files", how to run the game). Claude Code, Codex and Cursor all read it, so switching agents does not reset the project's rules. - Change one thing per round, run the game, and write down what changed. Small rounds are easy to undo.
1. Idea and a throwaway prototype
Decide what the player does in the first 30 seconds before anything else. Ask for the smallest playable version and play it yourself.
- one-sentence-to-game: one sentence to a small finished browser game, then one change per round with a log
- ccgs-brainstorm and ccgs-prototype: choose a concept, then a throwaway build with a proceed, pivot or kill verdict
Check: can you say the core loop in one sentence? Did you play the prototype for ten minutes and still want more?
2. Engine and code structure
Pick an engine whose project files are plain text, so the agent can read and diff everything. Godot 4 is a common choice for this reason.
- godot-nodes-scenes, godot-2d-movement and scenario-godot-2d for scene structure, controllers, tile maps and cameras
- save-systems as soon as there is progress worth keeping
- godot-prompter for a wider set of Godot skills
Check: the game runs from a clean checkout with one command, and the agent can explain where each system lives.
3. Art and animation
Fix the look first, then make assets that follow it. For animation, a video model can act out a move on a flat green background and a tool cuts the frames.
- ccgs-art-bible: write down palette, shapes and rules before mass-producing assets
- sprite-gen: sprite rows, chroma cleanup and atlases, plus the video-to-walk-cycle path (one seamless cycle, every direction starting on the same foot)
- anymove-creature-mocap: motion capture for 3D creatures from one AI clip (needs an NVIDIA GPU)
Check: every asset passes the art bible; walk loops have no seam or foot slide; nothing in the game copies an existing game's characters.
4. Sound
Placeholder sounds early make the game feel finished sooner. Replace them later.
- web-audio-sfx-synth: game sound effects from code, no files
- audio-design: buses, ducking, adaptive music and variation
Check: no sound repeats identically ten times in a row; music ducks under important cues; there is a mute.
5. Feel, balance and playtests
This is where hours go. Let the agent measure; you judge.
- game-feel: shake, hit-stop, easing and knockback
- ccgs-balance-check: outliers, broken progressions and degenerate strategies in your data
- ccgs-playtest-report: turn playtest notes into structured feedback
- performance-optimization and game-ui-ux: frame budget, menus and HUDs that work on every screen
Check: strangers finish the first session without you explaining anything; the slowest target machine holds its frame rate.
6. Build, trailer, store and launch
Automate the release so it is one command, not an afternoon: clean tree, build, upload, a changelog written from the commits.
- godot-export: export presets and headless builds for CI
- game-trailer-from-code: a 15-second trailer and the store images from your own running game
- steam-launch-check: review deadlines, refund-window risk from playtest times and the AI disclosure
- ccgs-launch-checklist: go or no-go across every department
Check: a fresh machine installs and runs the store build; the AI disclosure names what AI made; after launch, read every review and group the feedback before changing anything.
What the agent cannot do for you
It cannot tell you whether the game is fun, choose your audience, or carry the legal risk of assets that look like someone else's. Keep those decisions, and keep playing your own game every day.