The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted web content (DOM, scripts, assets) from user-provided URLs to perform shader extraction and reconstruction. * Ingestion points: Target webpage URLs, scripts, and asset definitions processed by scripts/fetch-rendered-dom.mjs and the recon kernel flow. * Boundary markers: The skill employs a strict state machine with defined gates (TARGET_LOCKED, REPLAY_READY) and structured JSON/Markdown templates to organize evidence and prevent bleed-through of untrusted data into instructions. * Capability inventory: The skill utilizes capabilities including runtime-eval, navigate, and network metadata access to analyze remote sites. * Sanitization: Instructions explicitly direct the agent not to persist sensitive session data like cookies or tokens, though the extracted code is intended for local reproduction.
[DYNAMIC_EXECUTION]: The skill instructions involve generating and executing code (JavaScript, GLSL, WGSL) to build local shader baselines and projectize the extracted visuals. * This dynamic generation is the primary purpose of the skill, facilitating the reconstruction of complex rendering pipelines. * Instructions provide templates for Three.js material modification and TSL (Three Shading Language) translation.

web-shader-extractor