The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes financial records, project notes, and attached documents which are externally controlled and could potentially contain malicious instructions.
Ingestion points: Data is ingested via project records, notes, and metadata through operations like listProjects, getProjectById, and attached documents as specified in SKILL.md.
Boundary markers: Present; the Boundaries section in SKILL.md explicitly instructs the agent to treat fetched records, notes, and documents as data to analyze, never as instructions.
Capability inventory: The skill can perform sensitive mutations, including creating, updating, and deleting financial periods and contract records via the buildr.execute tool.
Sanitization: The skill mandates a human-in-the-loop verification process, requiring a dry run and explicit user approval before any write operations are executed.
[COMMAND_EXECUTION]: The skill executes a local Python script to perform arithmetic operations on financial data.
Evidence: The workflow in SKILL.md directs the agent to run scripts/forecast_math.py with normalized JSON input.
Analysis: The script is bundled within the skill, uses only standard library modules, and performs no network calls or file writes. It includes input validation and handles errors gracefully, representing a secure implementation of local processing.

financial-forecasting