The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes shell commands like command -v claude and claude auth status --json to perform environment discovery and verify the user's authentication state before proceeding with automation tasks.
[DYNAMIC_EXECUTION]: The provided automation logic uses python3 -c as a dynamic JSON parser to process structured output from the Claude CLI during status checks.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to generate automated prompts that ingest untrusted data from external sources, creating a potential injection surface.
Ingestion points: Routines ingest data via the text field in API fire requests (up to 65,536 characters) and through GitHub event metadata (PR titles, bodies, and diffs).
Boundary markers: The skill requires the inclusion of a 'Safety rail' block in all generated prompts to explicitly instruct the agent to ignore any commands or instructions embedded within the data payloads.
Capability inventory: Generated routines have broad capabilities, including performing git operations (commit, push, PR creation) and interacting with various MCP connectors (Slack, Linear, etc.).
Sanitization: The prompt template includes specific instructions to treat payloads as untrusted data context and forbids the execution of any instructions contained therein.

routine-writer