The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on a custom command-line tool (subscription-research or digestpilot-runtime) and provides instructions for executing various Node.js scripts within the repository, such as scripts/doctor.mjs and scripts/bootstrap-runtime.mjs. These are used for workspace initialization and runtime setup.
[EXTERNAL_DOWNLOADS]: The content fetch command allows the agent to download full-text content from external URLs found in RSS feeds. This is a core feature for enriching research evidence but involves connecting to arbitrary third-party servers.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests untrusted data from external RSS feeds and websites.
Ingestion points: Untrusted content enters the agent's context through RSS summaries and full-text enrichment via subscription-research content fetch (stored in research.db and processed into briefs).
Boundary markers: The instructions emphasize treating CLI output as 'evidence packages' and separating synthesis from evidence, but the skill lacks technical delimiters (e.g., specific XML tags or markers) in its prompts to programmatically isolate this untrusted content.
Capability inventory: The agent can execute shell commands through the CLI and write synthesized markdown reports to the local file system (research-workspace/notes/daily/).
Sanitization: No specific sanitization or filtering logic for the fetched external text is described in the provided files.

subscription-research-agent