The Agent Skills Directory

[SAFE]: The skill implements a 'B-layer posture' which enforces a human-in-the-loop validation process. It uses specific sentinel strings ([AI-DRAFT], [MODELER INPUT NEEDED]) to ensure the AI does not finalize interpretive modeling claims or figure classifications without human ratification. This is a robust control for output integrity.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from various workspace files (JSON, Markdown, CSV). While this represents a theoretical injection surface, the skill lacks any execution capabilities (shell access, network operations, or file writing tools) that could be leveraged by a malicious payload embedded in the input data. The impact is limited to the skill's own Markdown output.
[SAFE]: All file path references are scoped to standard project directories such as workspace/, methods/, results/, and robustness/. No access to sensitive system paths, credentials, or environment variables is requested or performed.

figure-table-planner