knowledge/README.md

Abandoned path: We gave up on this as a from-scratch game-generation path. This knowledge base remains here as reference/source material for the pivot, especially for remixing known-good templates instead of generating games from nothing.

Knowledge base — game-patterns.txt

game-patterns.txt is the agent's game-development knowledge base: a code-forward reference of single-file HTML game patterns plus compact complete exemplars (game loop, canvas + DPR sizing, input, collision, physics, entities, HUD, tilemaps, audio, a minimal three.js boilerplate, GOTCHAS, and 10 playable genre examples). ~49 KB / ~12.3K tokens.

Where it came from

It was distilled from the vendored game-engine skill in ../skills/game-engine/, which is itself vendored from the github/awesome-copilot repository (credit and license noted in ../skills/README.md). We kept only the ~20% that helps a small model write single-file, no-build browser games (raw Canvas 2D + three.js via CDN importmap) and dropped the rest (publishing/marketing, glossary, architecture essays, Phaser/Babylon/A-Frame, Haxe, multi-file project tooling).

Complete compact exemplars

The Complete Compact Game Exemplars section gives the small model concrete games to imitate when patterns alone are not enough. Standalone copies live in examples/ so they can be opened and tested directly:

• top-down-racer.html
• pseudo-3d-racer.html
• canvas-platformer.html
• snake.html
• tetris.html
• sokoban.html
• raycaster-fps.html
• three-space-flight.html
• canvas-space-shooter.html
• three-obstacle-runner.html

Each exemplar is intentionally compact: one HTML file, inline CSS/JS, no build step, visible HUD, working controls, actual mechanics, and retry/win/lose behavior.

How the agent uses it (Traditional RAG)

agents.py embeds this file into a LanceDB vector store (db/lancedb/) with a local sentence-transformers embedder, and the agent runs Traditional RAG (search_knowledge=False, add_knowledge_to_context=True): the patterns relevant to each request are searched and injected straight into the prompt — the model never calls a search tool to fetch them.

Editing / regenerating

• Edit game-patterns.txt directly to change what the model gets.
• If an exemplar changes, update both game-patterns.txt and the matching file in examples/.
• After editing, delete db/lancedb/ so the patterns are re-embedded on next startup (insertion uses skip_if_exists=True, so unchanged content is not re-embedded).
• Keep examples compact — retrieved chunks share the model's context window with chat history and the game it has to generate.