The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data during multiple phases, which could contain malicious instructions.
Ingestion points: The skill reads external plan files (e.g., .claude/plan/xxx.md) and crawls project source code during the Phase 1 retrieval phase.
Boundary markers: The instructions include explicit boundary markers, requiring the agent to prioritize pre-approved plan steps and pause for user confirmation if the plan is missing or incomplete.
Capability inventory: The skill utilizes file modification tools (Edit/Write), sub-agent spawning (Agent tool), and shell command execution (Bash for validation).
Sanitization: Safety is managed via a mandatory human-in-the-loop confirmation phase (Phase 0) and local verification steps (Phase 3).
[COMMAND_EXECUTION]: The skill is instructed to execute shell commands for project verification.
Evidence: Phase 3 specifically instructs the agent to run lint, typecheck, and tests within the local environment to validate changes before delivery.

ac-execute