The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill sweeps external websites and search results to generate summaries and ranked findings. This creates a surface for indirect prompt injection where malicious content on a fetched page could attempt to influence the agent's summary or the next steps it proposes to the user.
Ingestion points: Web content is ingested via WebSearch and WebFetch tools as described in the Sweeper and Verifier briefs in BRIEFS.md.
Boundary markers: Absent; the external content is directly interpolated into the prompts for processing without specific delimiters or 'ignore embedded instructions' warnings.
Capability inventory: The skill has the ability to write files (findings and ledger files), perform network operations (web searches), and execute a local bash script.
Sanitization: No explicit sanitization of fetched text content is performed prior to processing, although the ledger management uses jq for structured data handling.
[COMMAND_EXECUTION]: The skill executes a local shell script scripts/filter-seen.sh to process and store a ledger of already-seen URLs.
Evidence: Steps 4 and 7 in SKILL.md invoke the script using bash to filter candidates and update the JSON ledger.
Context: The script is part of the skill's own directory and uses jq to handle JSON data safely, mitigating most common command injection risks.

chatter-scout