The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external business data (orders, invoices, and payments) from user-provided CSV and JSON files. This data is used to generate evidence chains and draft responses, such as chasing emails, for the AI agent to act upon. A maliciously crafted input file could attempt to inject instructions aimed at influencing the agent's behavior during follow-up tasks.
Ingestion points: scripts/import_tables.mjs (imports local data files into Busabase).
Boundary markers: The skill architecture forces a human review step; the agent is presented with a draft in a UI textarea that requires user approval or modification.
Capability inventory: Scripts like scripts/execute_decisions.mjs prepare operations for the agent, which are executed via other skills (e.g., kelly-email).
Sanitization: The skill performs data normalization and uses structured fields in Busabase, reducing the likelihood of successful injection through schema confusion.
[COMMAND_EXECUTION]: The skill includes several utility scripts (import_tables.mjs, run_checks.mjs, and execute_decisions.mjs) designed for the user to manage the audit workflow from the command line. These scripts perform file I/O and interact with the Busabase API.
[EXTERNAL_DOWNLOADS]: The skill and its included AirApp components interact with the Busabase platform (busabase.com) to synchronize data. This interaction is central to the skill's purpose and uses documented OAuth and API protocols.

kelly-audit