The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes shell commands to extract repository metadata, such as using git shortlog and git log to infer team membership and project age. It also executes repository-specific verification commands (e.g., test runners and linters) defined in the configuration to validate the setup. Additionally, it creates a shell script in the .claude/hooks/ directory and applies chmod +x to make it executable.
[PROMPT_INJECTION]: The skill implements a persistence mechanism by offering to modify the agent's configuration file (.claude/settings.json) to add a SessionStart hook. This hook triggers the automatic execution of a script every time the agent session is initiated or reset, which can persistently influence agent behavior within that environment.
[PROMPT_INJECTION]: The skill ingests data from various repository files to generate configuration drafts, creating a surface for indirect prompt injection.
Ingestion points: Manifests and metadata files including package.json, Cargo.toml, pyproject.toml, Makefile, CODEOWNERS, AUTHORS, and CONTRIBUTORS.
Boundary markers: No explicit markers or delimiters are used to wrap the ingested content.
Capability inventory: The skill has the capability to write files to the docs/agents/ directory, modify agent settings, and execute arbitrary commands defined as 'verify' steps.
Sanitization: The risk is mitigated by a mandatory human-in-the-loop step, where the user must review and explicitly confirm all drafted content before it is written or executed.
[EXTERNAL_DOWNLOADS]: The skill recommends the use of an external service, the 'Context7 MCP server', for library documentation, though it does not perform the installation automatically.

setup-repo