The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary function involves scanning the active session transcript to identify key decisions and state, which is then written to local files (.agents/anchor.md). This creates a surface where external or untrusted data from the session history could be persisted and later read by the agent. This is mitigated by the following factors:
Ingestion points: Active session history scanned during the 'Scan Active Session' step.
Boundary markers: The skill uses a structured Markdown format with specific headers (workstream:, branch:, Client:) to delimit data.
Capability inventory: The skill only utilizes view_file and write_to_file tools, lacking dangerous capabilities like network access or shell execution.
Sanitization: The instructions explicitly include a 'Client-Confidentiality Guard' and a 'No secrets' rule, directing the agent to never include tokens, credentials, or .env values in anchors.
[SAFE]: The installation command referenced in the documentation (npx skills add harshsinghmp/muse-skills) is a user-facing instruction for skill management and does not involve the agent executing remote code during runtime. The resources originate from the skill's author.

context-anchor