The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches several specialized CLI tools and plugins from the NPM registry and GitHub to support its development workflows. This includes @earendil-works/pi-agent-browser for UI verification, @vedanth/context7 for documentation search, and npx skills from Vercel Labs for loading coding patterns. These downloads are standard for the tool's operational requirements.
[COMMAND_EXECUTION]: The orchestrator dynamically executes verifyCommands (such as unit tests or linters) that it extracts from generated technical plans. This behavior is gated by mandatory human approval of the technical plan (spec-tech.md) before any execution occurs.
[PROMPT_INJECTION]: The skill exhibits a surface for indirect prompt injection as it processes external project specifications (spec-product.md) to generate execution plans. This risk is mitigated by multiple security controls:
Ingestion points: Technical context and product specs are read from the .stelow/ workflow directory.
Boundary markers: Mandatory visual review gates using the plannotator tool are enforced at the Shape, Interface, and Planning stages to ensure human oversight.
Capability inventory: The agent has access to shell execution, subagent delegation, and network operations via specialized tools.
Sanitization: The skill performs structural validation on generated scopes, checking for required fields (TYPE, DoD, AC) and circular dependencies before execution.
[COMMAND_EXECUTION]: Shell scripts within the skill use node -e to execute small JavaScript snippets for parsing and filtering JSON configuration from stelow.json. This is used as a lightweight alternative to dedicated JSON processors like jq in the environment.

stelow-product-tech-planning