The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill acts as an orchestrator that ingests external, untrusted content such as Product Requirement Documents (PRDs), design specifications, and codebase files during the 'start' and 'validate' phases. This information is used to generate implementation plans, architectural decisions, and 'smoke test' verification commands. This creates an attack surface where malicious instructions hidden in documentation could influence the plan's tasks or the verification scripts executed by the agent.
Ingestion points: start/SKILL.md (Phase 1: PRD & Design Analysis) and validate/SKILL.md (reading the generated plan file).
Boundary markers: The skill implements several 'gates' (CEO review, Engineering review, specialist validation) and explicitly includes checkpoints for user confirmation before proceeding to planning or execution.
Capability inventory: The skill has access to Bash (shell execution), Write (file modification), and Task (sub-agent spawning).
Sanitization: No explicit string sanitization is documented; however, the use of specialist security-analyst and security-reviewer agents is designed to mitigate this risk.
[COMMAND_EXECUTION]: The execute sub-skill uses the Bash tool to perform Git operations (worktree management, commits, pushing) and GitHub CLI commands (gh pr create). Crucially, it executes 'smoke test' commands defined in the plan's ## Integration Verification section, which are generated based on the initial planning phase. This represents execution of commands derived from processed data.
[DYNAMIC_EXECUTION]: The pipeline utilizes the Task tool to spawn multiple specialist sub-agents (e.g., code-explorer, security-analyst, security-reviewer) to perform research and validation in parallel. The instructions for these agents are dynamically constructed from the plan and research results.

plan-pipeline