The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests data from external sources, such as competitor websites and market trend movers, which creates a potential surface for indirect prompt injection attacks.
Ingestion points: External data enters the system through scripts/ingest_signals.mjs (competitor monitoring), scripts/ingest_trends.mjs (search and community trends), and scripts/file_report.mjs (research findings).
Boundary markers: The skill enforces a mandatory human triage step within the Kelly Radar AirApp. The agent is instructed that all handoffs and outbound artifacts require explicit approval.
Capability inventory: The skill can write to Busabase, propose handoffs to sibling skills (e.g., kelly-writer), and print execution plans for the agent to follow.
Sanitization: Ingestion scripts include schema validation, content hashing for deduplication, and instructions to redact personal data from signals and reports. The human review requirement effectively mitigates the risk of autonomous execution of injected instructions.
[COMMAND_EXECUTION]: The skill includes several utility scripts (e.g., setup.mjs, publish_airapp.mjs) used for provisioning the workspace and deploying the AirApp. These operations are part of the intended platform functionality for Busabase App-in-Skills and are executed by the trusted user rather than automatically by the agent at runtime.
[CREDENTIALS_SAFE]: Secret management follows best practices. API keys and credentials for Busabase and external crawling services are accessed via environment variables (BUSABASE_API_KEY, BUSABASE_SPACE_ID) rather than being hardcoded or stored in the database.

kelly-radar