The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it performs a deep audit of untrusted game files.\n
Ingestion points: Step 1 involves reading the entire game codebase, including package.json, .js files, index.html, and tests/ (found in SKILL.md).\n
Boundary markers: There are no explicit instructions to the agent to ignore embedded instructions or use delimiters when reading these files.\n
Capability inventory: The skill can modify the filesystem (Step 4) and execute shell commands through npm (Steps 4 and 5).\n
Sanitization: No sanitization or validation of the content of the read files is performed before processing them as instructions or data.\n- [COMMAND_EXECUTION]: The skill executes local shell commands to build and test the game.\n
Evidence: The instructions in Step 4 and Step 5 specify running npm run build, npm test, and npm run test:update-snapshots. These commands execute scripts defined in the project's package.json, which could be used to run arbitrary code if the project files are malicious.\n- [DATA_EXFILTRATION]: The skill accesses a wide range of local project files which may contain sensitive configuration or structure.\n
Evidence: Step 1 explicitly instructs the agent to read package.json, configuration values in Constants.js, and the entire directory structure for scenes, entities, and UI.

improve-game