The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed specifically to process untrusted student materials. It mitigates the threat of indirect prompt injection using a dynamic data fencing system (<<<MATERIAL <id> ... MATERIAL>>> <id>) where the identifier is a SHA256-derived hash of the content, preventing data from escaping into instructions. Evidence Chain: (1) Ingestion points: coach/extract.py reads user-specified document folders. (2) Boundary markers: Present and dynamically verified via content hashing. (3) Capability inventory: Local file writing for workspace management, progress logs, and Markdown notes; no network or administrative capabilities. (4) Sanitization: coach/guard.py implements regex-based filters to detect AI instructions and specifically removes zero-width Unicode characters.
[COMMAND_EXECUTION]: Local Python scripts (coach.py, coach/cli.py) are executed to manage study state and parse documents. These operations are restricted to the skill's workspace and are fundamental to its stated purpose.
[EXTERNAL_DOWNLOADS]: Recommends the installation of standard, well-known libraries pypdfium2 or pypdf for PDF processing; no unauthorized or hidden downloads were found.
[OBFUSCATION]: Static analysis flagged defensive code in coach/guard.py as suspicious; however, this code is a security feature that normalizes text and removes hidden Unicode characters from source materials to prevent steganographic prompt injection.

universal-exam-cram-coach