The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill includes a shell script, scripts/probe_asset_credentials.sh, which is specifically designed to probe for environment variables containing API keys (TRIPO_API_KEY, GEMINI_API_KEY, ELEVENLABS_API_KEY). The skill instructs the agent to run this script and report whether keys are SET or MISSING. Accessing sensitive shell profile files to search for credentials is a high-risk activity.
[DYNAMIC_EXECUTION]: The scripts/probe_asset_credentials.sh script executes the contents of sensitive user shell configuration files (~/.bashrc, ~/.zshrc, ~/.bash_profile, ~/.zprofile) using the source command (or . in zsh). This triggers the execution of any arbitrary code defined in those files within the script's execution environment.
[COMMAND_EXECUTION]: The skill relies on several shell-based tools and scripts for its core operations, including scripts/probe_asset_credentials.sh for credential checking, scripts/audit_reference_report.py for report validation, and instructions for running node and python3 for project scaffolding and canvas inspection.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process a wide variety of external data, including sibling skill definitions (SKILL.md files), markdown reports, and user-provided game design documents. Because the skill possesses significant capabilities (shell execution, file system access), it is vulnerable to malicious instructions embedded in these processed files.
Ingestion points: Reads sibling SKILL.md files from relative or home directory paths, and reads project reports for auditing in scripts/audit_reference_report.py.
Boundary markers: No explicit delimiters or boundary markers are defined to separate untrusted data from instructions when the agent processes external skill files or reports.
Capability inventory: Provides shell command execution via bash, python3, and node, as well as file system read/write access.
Sanitization: No evidence of sanitization, validation, or filtering of the content found in sibling skills or audited reports before the agent acts on them.

threejs-game-director