The Agent Skills Directory

Indirect Prompt Injection Defenses: The skill processes external plans and evidence (ingestion points in SKILL.md) which could contain untrusted instructions. It explicitly includes boundary markers by instructing the agent to treat these inputs as data and ignore any embedded directives. The skill's capabilities are limited to reading these files, dispatching subagents for analysis, and writing a single output artifact (capability inventory in SKILL.md). The presence of these specific sanitization instructions (sanitization in SKILL.md) reduces the risk of successful indirect prompt injection.
Structured Subagent Dispatch: The skill utilizes a parallelized 'lens' assessment approach, dispatching specific tasks to a subagent. This process is governed by a defined reasoning effort protocol and operates within the host environment's standard toolset, posing no unusual security concerns.
Least Privilege File Operations: The skill is restricted to a read-only assessment of the plan and is permitted to write only a single, caller-specified critique artifact. These constraints ensure that the skill operates within a well-defined boundary without modifying existing project files.

rpi-plan-critique