The Agent Skills Directory

[COMMAND_EXECUTION]: Several bundled utility scripts, including scripts/run_production_evals.py, scripts/validate_package.py, and scripts/test_eval_validators.py, utilize subprocess.run() to execute system commands such as git archive, node, and python3. These operations are used to manage the evaluation environment and run structural validators against generated Markdown files. The commands generally target local paths within the skill package or an isolated workspace.
[DYNAMIC_EXECUTION]: The script scripts/validate_package.py uses the Python compile() function to perform syntax checks on other scripts within the package. This is used for linting purposes to ensure package integrity and does not involve the execution of the resulting code via eval() or exec().
[INDIRECT_PROMPT_INJECTION]: The skill defines a significant attack surface by instructing the agent to ingest external data from digital storefront documentation (e.g., Steamworks, Google Play developer console) via web tools in references/document-types/design-system/storefront-research.md. The agent then processes this data to generate design documents while possessing capabilities such as file system writing and shell access.
Ingestion points: Web search and fetch operations triggered by storefront research instructions; reading existing user-supplied documents for normalization or updates.
Boundary markers: Present; the skill uses rigid Markdown templates (FDD, Design System), numbered section requirements, and metadata blocks that help isolate intended content from potentially malicious data.
Capability inventory: The skill configuration allows reading and writing to the workspace, web access, and the execution of bundled validation scripts via Bash.
Sanitization: Absent; the skill relies on the agent's instructions to faithfully reproduce and normalize source facts rather than filtering external inputs.

document-writing