The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input from design documents (e.g., docs/design/intent-confirmation-domain-design.md) to generate milestone specifications. This ingestion of external data represents a potential indirect prompt injection surface. However, the skill provides robust mitigation through a mandatory five-section template and strict 'DoD Assertion Writing Rules' that explicitly forbid prose and require machine-verifiable assertions. These constraints effectively limit the influence of potentially malicious instructions embedded in the input data.\n- [COMMAND_EXECUTION]: The skill references various shell commands such as npx vitest run, tsc, and playwright within its instructions. These are provided as examples of verifiable assertions that an agent should write into the generated documentation. The skill does not execute these commands directly; instead, it defines how they should be recorded for later verification by a downstream 'stage-gate' tool. The skill relies on an external skill, project-handoff, for handling file storage interactions in the .handoff/ directory, maintaining a clear separation of concerns and minimizing direct command execution risk.

stage-spec