The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill reads and follows instructions from untrusted project files like CLAUDE.md to determine directory preferences and executes scripts defined in project configuration files.
Ingestion points: Reads CLAUDE.md via grep for worktree directory preferences; detects package.json, Cargo.toml, requirements.txt, pyproject.toml, and go.mod to trigger setup and test commands.
Boundary markers: None identified; instructions extracted from CLAUDE.md or inferred from project files are processed without explicit delimiters or safety warnings.
Capability inventory: Executes shell commands for package installation (npm install, cargo build, pip install, poetry install, go mod download), runs test suites (npm test, cargo test, pytest, go test ./...), and performs Git operations including git commit and git worktree add.
Sanitization: None; the agent is instructed to use preferences found in CLAUDE.md "without asking" and to "auto-detect and run" setup commands based on file presence.
[COMMAND_EXECUTION]: The skill automatically executes setup and test commands. This involves running code defined within the project's metadata (such as npm scripts or test suites), which is a common practice in development but technically executes arbitrary code from the current workspace.
[SAFE]: The use of standard package managers (npm, pip, cargo, poetry, go) to fetch dependencies from well-known registries is a routine development task and is considered safe behavior for a development-oriented skill.

using-git-worktrees