The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a recursive loop where 'builder' and 'judge' agents process data from paprika-loop/PROMPT.md and paprika-loop/LATEST.md.
Ingestion points: paprika-loop/PROMPT.md (user-editable brief) and paprika-loop/LATEST.md (output from previous AI sessions) are read at the start of every loop iteration.
Boundary markers: The instructions define these files as the source of truth for the request but do not specify technical delimiters or "ignore" instructions to prevent injection within the data.
Capability inventory: The skill uses the Write tool to modify files, executes CLI commands (node, bun, git, claude), and performs network-based asset searches.
Sanitization: No sanitization or validation of the brief's content or the previous session's output is performed before processing.
[COMMAND_EXECUTION]: The skill initiates background processes to manage the loop lifecycle. The agent is instructed to execute node ${CLAUDE_PLUGIN_ROOT}/src/cli.js loop or bun run ${CLAUDE_PLUGIN_ROOT}/src/cli.ts loop from the project folder.
[DYNAMIC_EXECUTION]: The skill generates and executes code as part of its core functionality. The 'builder' session writes Three.js code (index.html, geo.js, fx.js) which is subsequently rendered and evaluated by the 'judge' session using tools like open and look.

paprika-loop