The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from RSS feeds, competitor posts, and Threads discussions to generate content insights.
Ingestion points: Untrusted external data enters the agent context through the evidence, threads, and research:add commands.
Boundary markers: The 'Analyst brief' provided in SKILL.md includes explicit instructions for subagents to ignore any instructions embedded in third-party text.
Capability inventory: The skill uses a bundled script for network communication with a backend service and performs local file writes to manage research batches.
Sanitization: Safety is managed by delegating analysis to subagents with strict guidelines to avoid following embedded instructions.
[EXTERNAL_DOWNLOADS]: The skill makes network requests to an external API endpoint hosted on a cloud platform (intent-crab-142.convex.site) to synchronize campaign data and upload media.
[COMMAND_EXECUTION]: The skill executes a bundled helper script (./scripts/bichon.cjs) to interact with its API. The script implements security best practices such as redacting API keys from console output and applying restrictive (0600) permissions to local configuration files.

bichon