The Agent Skills Directory

Indirect Prompt Injection Surface: The skill processes data from various external sources, including connected accounting platforms and user-uploaded CSV or XLSX files.
Ingestion points: Data enters the context via tool outputs from financial connectors and by reading user-uploaded files, as described in SKILL.md and reference/data_sources.md.
Capability inventory: The agent utilizes Read and WebFetch tools, writes to the filesystem, and executes generated scripts to create report workbooks.
Sanitization & Boundaries: The instructions do not detail explicit sanitization or delimiters for external data. If an external source contains adversarial content, it might influence the agent's behavior during report generation.
Dynamic Script Execution for Artifact Generation: The agent is instructed to generate and execute scripts to build the final XLSX workbooks rather than assembling them manually.
Description: The agent creates code at runtime to handle data processing and document formatting tasks.
Security Consideration: Executing dynamically generated code is a notable pattern, particularly when the logic may be influenced by data originating from untrusted external sources.
State Persistence via Local Configuration Files: The skill maintains recurring report definitions by writing to and reading from local files like report-definitions.md.
Description: This allows the skill to maintain persistence for the user's report specifications across different sessions.
Security Consideration: The integrity of future report runs depends on the security of these locally stored files, which the agent is instructed to trust for automated re-execution.

report-builder