The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes user-supplied input through the $ARGUMENTS variable to determine project structure and execution flow without sanitization.
Ingestion points: $ARGUMENTS variable in SKILL.md frontmatter and body.
Boundary markers: Absent. No instructions are provided to the agent to validate or escape the input before using it in directory creation or verification steps.
Capability inventory: The skill has the capability to write files, initialize git repositories, and execute build/lint commands (subprocess calls).
Sanitization: Absent. A malicious project name or type could lead to command injection if the agent interpolates the input directly into shell commands during the scaffolding or verification process.
[COMMAND_EXECUTION]: The skill generates .claude/settings.json configuration files that include automated shell hooks (PostToolUse and Stop) designed to execute on every file write.
Evidence: The Python and Next.js templates generate hooks that run shell pipelines (e.g., python3 -c, ruff, npx biome) whenever a file is edited.
Context: While intended for auto-formatting, this establishes a mechanism for persistent, automated command execution within the local development environment.
[REMOTE_CODE_EXECUTION]: The skill explicitly instructs the agent to run build and lint verification steps that involve external package managers and toolchains.
Evidence: Execution Step 7 in SKILL.md requires the agent to "check that the project builds/lints after scaffolding."
Context: This results in the execution of commands like npm run build, uv run python, and flutter run, which download and execute code from remote registries (npm, PyPI, etc.).

project-template