The Agent Skills Directory

[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface when processing human feedback. The user-provided <feedback> is interpolated into the context for the /brainstorm skill call without explicit sanitization or boundary markers.
Ingestion points: The feedback argument passed to the skill in Human Feedback mode.
Boundary markers: Absent; the feedback is passed as a string context to subsequent skills.
Capability inventory: The skill can invoke /brainstorm, /investigate, /formalize-problem, and /synthesize, which triggers further file reads and command executions.
Sanitization: None mentioned; the feedback is used verbatim or lightly paraphrased.
[DATA_EXFILTRATION]: In --codex mode, the skill extracts segments of research data, including the last five findings from current-understanding.md and summary rows from notes/results.md, and sends them to an external OpenAI Codex instance via MCP. While this is the primary functionality of the consultation mode, it involves exporting internal workspace data to a third-party service.
[COMMAND_EXECUTION]: The skill orchestrates the execution of multiple other agent skills (/brainstorm, /investigate, /formalize-problem, /synthesize) based on the classification of feedback or Codex consultation results. This creates a chain of automated actions triggered by external input.
[OTHER]: The skill includes a 'Path Resolution Protocol' that instructs the agent to manually resolve paths by substituting {{REAPER_SKILL_DIR}} with absolute directories such as ~/.claude/skills/reaper/ or ~/.agents/skills/reaper/. This practice guides the agent to access specific configuration and reference files outside its immediate workspace.

critique