The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill bundle features five components (Comment Drafter, Reply Handler, Hook Extractor, Thread Monitor, and Engager Analytics) that ingest untrusted content from LinkedIn, such as post bodies and comment threads.
Ingestion points: Data enters through lib/apify_client.py using actors to fetch remote content from public LinkedIn posts.
Boundary markers: The skill incorporates a robust security policy in references/untrusted-content.md. It explicitly instructs the agent to treat fetched data as non-authoritative and ignore instructions found within it, effectively delimiting data from instructions.
Capability inventory: The agent can publish content via API or local command execution after user approval.
Sanitization: Content is filtered for identifiers and URNs using regex in lib/url_parser.py rather than being passed raw to sensitive operations.
[DYNAMIC_EXECUTION]: The skill bundle includes an advanced 'DIY' backend that executes shell commands defined in a local environment variable.
Evidence: In lib/backend_selector.py, the skill uses subprocess.run to invoke a command string retrieved from the LINKEDIN_SKILLS_CUSTOM_POSTER environment variable.
Context: This is a documented extensibility feature designed for advanced users to integrate custom posting tools. It uses shlex.split for safe argument parsing and executes without a shell to minimize injection risks. It is only triggered after explicit user approval of a draft.

linkedin-marketing