The Agent Skills Directory

[SAFE]: The skill implements a robust human-in-the-loop mechanism ('Step 4: Confirm before writing') that prevents silent or automated sharing of data. The agent must display the intended target and content to the user and wait for explicit approval before proceeding with the write_note operation.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an indirect prompt injection surface because it reads note content which could contain malicious instructions and then processes that content to be written elsewhere.
Ingestion points: Data is ingested through the $ARGUMENTS permalink/title and the resulting content from the read_note tool (SKILL.md, Step 2).
Boundary markers: The skill uses boundary markers by adding specific frontmatter (shared_from) and context observations to the shared note (SKILL.md, Step 5).
Capability inventory: The skill uses the write_note tool to publish content to team projects (SKILL.md, Step 5).
Sanitization: No automated sanitization is described, but the mandatory user approval step serves as a significant manual validation checkpoint.
[SAFE]: Configuration handling for CLAUDE_CONFIG_DIR and settings.json is described with literal resolution logic that avoids dangerous behaviors like automatic tilde expansion or trimming, which helps prevent unintended path resolution.

bm-share