The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to run an automated shell command (ls) targeting several hidden directories in the user's home folder (e.g., ~/.claude/skills/, ~/.agent/skills/) to identify other installed companion skills upon first activation.
[PROMPT_INJECTION]: Employs highly assertive behavioral overrides using 'HARD-GATE' and 'MUST' directives to strictly enforce a multi-phase workflow and ensure the agent stays in 'plan mode' until design approval.
[PROMPT_INJECTION]: The skill presents a significant surface for Indirect Prompt Injection by instructing the agent to ingest and reason over large volumes of untrusted data from the local codebase.
Ingestion points: Phase 1 ('Deep context scan') reads project files such as README.md, CLAUDE.md, and all files in the docs/ directory. Phase 2 ('Codebase-First Intelligence') uses grep and glob to scan the entire codebase for context before asking questions.
Boundary markers: Absent. The instructions do not define delimiters (like XML tags or markdown blocks) or provide 'ignore embedded instructions' warnings for the data being read.
Capability inventory: The agent possesses file system read/write capabilities (writing design specs to docs/plans/), shell execution (ls, grep, glob), and the ability to trigger subagents for 'Spec Review' loops.
Sanitization: Absent. There is no evidence of content filtering or validation for the data retrieved from the codebase before it is processed by the agent.

super-brainstorm