The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-provided inputs to generate technical specifications, which represents a potential attack surface, though it includes significant mitigations.
Ingestion points: Reads a 'user ask' or 'ticket-ref' to initiate the alignment process (SKILL.md).
Boundary markers: Utilizes a structured 'Decisions' log with dated Q&A entries and follows a rigorous specification template (references/spec.md).
Capability inventory: Includes file writing, shell command execution (git, pnpm), and subagent/MCP tool invocation.
Sanitization: Explicitly mandates that 'Facts come from subagents, never from the user' and requires all measured numbers to be verified against production targets or artifacts (SKILL.md, references/spec.md).
[COMMAND_EXECUTION]: The skill executes standard development and testing commands as part of its primary software planning and verification purpose.
Evidence: Uses pnpm test, pnpm dev, and git commands for testing stubs and managing repository state (SKILL.md, references/spec.md).
[DYNAMIC_CONTEXT_INJECTION]: The skill text includes instances of shell command execution syntax used for administrative tasks.
Evidence: Mentions ! git rm <file> as a method for clearing stale locks during the alignment process (SKILL.md).
[DATA_EXFILTRATION]: Contains robust, explicit prohibitions against the exposure of credentials or sensitive configuration data.
Evidence: Strictly forbids opening .env, .mcp.json, or other configuration files holding credentials, and ensures API keys are never logged (SKILL.md, references/spec.md).

align