The Agent Skills Directory

[SAFE]: The skill implements protective instructions ('Invariants') that explicitly forbid the agent from inspecting .gitignore files or committing code without explicit user consent. This prevents the accidental exposure of sensitive local credentials and ensures the user retains control over the repository state.
[COMMAND_EXECUTION]: The skill writes generated documentation to a local docs/agentic-engineering/ subdirectory. These operations are limited to specific, documented paths and do not involve executing arbitrary shell scripts or downloading remote content.
[PROMPT_INJECTION]: The skill has an indirect prompt injection surface as it ingests project data (Stage 1: 'Check out the current project state first') and processes it into documentation.
Ingestion points: Stage 1 Understanding reads repository files, documentation, and commit history.
Boundary markers: Absent; the skill does not use specific delimiters or instructions to ignore embedded commands in the analyzed files.
Capability inventory: The skill performs file system writes (docs/agentic-engineering/) but lacks high-risk capabilities like network access, eval/exec, or system-level configuration changes in its instructions.
Sanitization: Absent; content from the repository is interpolated directly into templates.

brainstorm