[P] Visual verification as a feedback loop for LLM code generation

A developer has created an open-source, autonomous pipeline that generates playable Godot games directly from text prompts, tackling the dual challenges of LLMs writing correct code in underrepresented languages and verifying that correctness beyond mere compilation. The system specifically addresses GDScript—Godot's Python-like scripting language with ~850 classes—which has limited representation in standard LLM training data. To overcome this, the pipeline implements a sophisticated three-layer reference system: a hand-written language specification detailing GDScript's unique behaviors, full API documentation for all engine classes converted to compact Markdown, and a database of engine quirks that aren't apparent from documentation alone.

The core innovation lies in its agentic lazy-loading approach to context management. Since loading all 850+ class docs would consume the entire context window, the system uses a two-tier index where a small set of common classes is always loaded, while the agent dynamically loads specific class documentation only when needed. This ensures the LLM (Claude Code) can access necessary references without being overwhelmed. Verification occurs in three stages: compilation checking via Godot's headless mode, agentic screenshot analysis where the coding agent captures and assesses running scenes, and visual validation that elements render correctly. The entire process runs in forked contexts to prevent state degradation, making context selection—not just code generation—the critical success factor.