The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute a local Python script ../obsidian-notes-karpathy/scripts/scan_compile_delta.py to identify changes in the raw data directory. This is a standard part of the incremental compilation workflow.
[PROMPT_INJECTION]: The skill ingests untrusted data from markdown and PDF files located in raw/ subdirectories. An attacker could embed instructions in these files to influence the summarization or classification behavior.
Ingestion points: Reads markdown and PDF content from raw/human/**, raw/agents/{role}/**, raw/*.md, and raw/**/papers/*.pdf.
Boundary markers: The instructions do not define explicit delimiters or warnings for the agent to ignore embedded instructions within the raw content.
Capability inventory: The skill has the ability to write files to wiki/drafts/ and wiki/log.md, and execute the local scan_compile_delta.py script.
Sanitization: There is no explicit sanitization step, but the skill requires clear separation between source claims and compiler inferences in the output, which provides some structural mitigation.

kb-compile