The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests external project documentation and user-provided feature descriptions to drive its automated workflow, creating a potential surface for indirect instructions.
Ingestion points: Processes user input in featureDescription and reads existing project metadata from files like AGENTS.md or skills/<projectSlug>/SKILL.md.
Boundary markers: The instructions do not define explicit delimiters to separate untrusted project data from internal logic.
Capability inventory: The skill can write new executable skill files, modify source code, and execute shell commands for testing and builds.
Sanitization: Risks are mitigated by 'Socratic Discovery' and 'Plan Approval' gates, which require explicit human confirmation before the agent proceeds to execution phases.
[DYNAMIC_EXECUTION]: The workflow involves the creation of new executable logic and the execution of generated plans.
The skill scaffolds new executable 'workflow contracts' (SKILL.md files) in the skills/ directory during Phase 5, effectively generating new agent behaviors at runtime.
It performs automated updates to existing project skills and routing matrices in Phase 10 based on discoveries made during the development cycle.
[COMMAND_EXECUTION]: The skill invokes external development tools and environment-specific commands to verify implementation quality.
Phase 7 and Phase 9 involve running project test suites and build verification commands, including bun run build, npm run test, rspec-rails, vitest, and pytest.
Infrastructure-related tasks include pre-change checks and service health validation, which may involve network or system-level queries depending on the project environment.

radioactive