The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data (plan files or project briefs) by interpolating the content directly into a subagent prompt for multi-lens review.
Ingestion points: Untrusted data enters the agent context via the <plan file or brief> placeholder in the subagent prompt skeleton defined in SKILL.md.
Boundary markers: The skill does not define specific delimiters or "ignore embedded instructions" warnings to wrap the untrusted input within the subagent prompt.
Capability inventory: The skill possesses Bash, Write, Edit, and Agent tool access, which represents a capability tier that could be targeted by instructions hidden in user-provided documents.
Sanitization: There is no mention of sanitization, escaping, or validation of the plan content before it is processed by the agent.
[COMMAND_EXECUTION]: The skill explicitly instructs the agent to use the Bash tool to perform file integrity checks after operations.
Evidence: The Gotchas section in SKILL.md (dated 2026-08-06) directs the use of wc -l <file> and tail -c 60 <file> to detect and mitigate silent file truncation issues after Write operations.

planning-pipeline