The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted data by reading weekly digests and identifying source markers (SKILL.md, Workflow Steps 1-4). While boundary markers for these instructions are absent, the skill's capabilities—which include network reading (WebFetch) and file writing to the project's watchlist—are mitigated by a mandatory human-in-the-loop review. The workflow requires the agent to emit a proposed diff and wait for explicit user approval before applying changes, preventing automated exploitation by external content.\n- [SAFE]: The skill follows defensive design principles by establishing a 'propose-then-approve' pattern. It includes specific guardrails such as capping the number of monthly additions, protecting 'essential' sources from accidental deletion, and providing a clear audit trail through changelog entries. No malicious code, obfuscation, or persistence mechanisms were detected.

update-watchlist