The Agent Skills Directory

[PROMPT_INJECTION]: The skill defines an 'Autonomous Execution Policy' that explicitly instructs the agent to never pause, stop, or ask the user questions. This suppresses the human-in-the-loop oversight standard for safe agent operations.- [INDIRECT_PROMPT_INJECTION]: The skill ingests data from external project files (PRDs, TechSpecs, task definitions) to drive implementation and testing tasks without safety boundaries or sanitization.
Ingestion points: Files located at ./prds/prd-[feature-slug]/tasks/[num]_task.md, ./prds/prd-[feature-slug]/prd.md, and project source code.
Boundary markers: Absent. The skill does not use delimiters or instructions to ignore potential commands embedded in these files.
Capability inventory: The skill has broad capabilities including Write and Edit tool access, git command execution, and the ability to run arbitrary tests and dev servers via MCP tools.
Sanitization: Absent. The skill is instructed to treat all tags and requirements in these files as mandatory constraints.- [COMMAND_EXECUTION]: The skill performs environment checks (scanning for .claude/, .github/, .cursor/), identifies package managers, and executes project scripts (e.g., npm run dev, git rev-parse) to perform its duties.- [DYNAMIC_EXECUTION]: The skill dynamically resolves its execution environment and then loads and follows instruction sets (skills) from computed paths such as .claude/skills/ or .cursor/rules/. It also dynamically maps MCP server capabilities at runtime.

do-execute-task