The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's core workflow relies on agents reading and following state, insights, and mapping instructions from files like WORKSPACE.md and MEMORY.md. This creates a vulnerability where malicious content within these files could override agent behavior in subsequent sessions.
Ingestion points: WORKSPACE.md, MEMORY.md, and neighboring WORKSPACE.md files (SKILL.md lines 50, 71).
Boundary markers: None specified; agents are instructed to treat these files as a source of authority.
Capability inventory: File system read/write access is implied for managing research artifacts and updating workspace maps.
Sanitization: No validation or sanitization of the content within the mapped research files is mentioned.
[DATA_EXFILTRATION]: (Data Exposure) The skill explicitly encourages agents to access and map directories outside the immediate project root, which could lead to unintended exposure of sensitive data if used in a broad file system context.
Evidence: The WORKSPACE.md specification supports relative paths like ../neighbor-project/research/notes/ for cross-linkage, and the 'Smart Discovery' heuristics instruct agents to establish research continuity by reading WORKSPACE.md files in neighboring directories (SKILL.md lines 40, 50).

research-workspace-standard