The Agent Skills Directory

[DATA_EXFILTRATION]: The skill accesses and writes to ~/.claude/projects/skill-telemetry.md. This hidden path in the user's home directory is used to persist execution metadata across sessions, which serves as a persistence mechanism and involves accessing sensitive, user-specific application data.
[COMMAND_EXECUTION]: The 'Build Verification' and 'Self-Healing Validation' phases require the agent to run engine-specific build tools and project test suites. Executing these commands on an untrusted codebase—which may contain malicious build scripts, configuration files (e.g., package.json scripts), or test code—results in arbitrary command execution on the host system.
[PROMPT_INJECTION]: The skill is highly susceptible to indirect prompt injection because its primary function is to ingest and analyze entire game projects. Malicious instructions embedded in source code comments or string literals could override the agent's logic during the audit.
Ingestion points: Source files (*.cs, *.cpp, *.gd, *.ts, *.js), project manifests (package.json, project.godot), and scene files are processed for analysis.
Boundary markers: Absent. The skill does not utilize delimiters or specific instructions to ignore embedded commands within the analyzed data.
Capability inventory: The skill performs file system read/write operations and executes shell commands via build and test utilities.
Sanitization: Absent. The external content from the codebase is processed directly without escaping or validation before being used to inform 'self-healing' fixes or analysis reports.

game-qa