The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The adapters/codex/scripts/llm_wiki_ingest_arxiv.py script fetches paper metadata and abstracts from the official arXiv API (export.arxiv.org). This activity is a well-documented and expected feature for automating research workflows.
[COMMAND_EXECUTION]: The skill uses Bash and exec_command tools to run internal Python scripts (llm_wiki_lint.py, llm_wiki_scaffold.py, etc.) for wiki maintenance. These scripts are provided within the skill's own directory and operate strictly on the user's defined wiki folders.
[PROMPT_INJECTION]: The skill has an inherent surface for indirect prompt injection because it ingests and processes untrusted markdown content from external URLs and PDFs. Crafty external content could include instructions designed to influence the agent's behavior during synthesis.
Ingestion points: SKILL.md (Ingest operation) and adapters/codex/scripts/llm_wiki_ingest_arxiv.py.
Boundary markers: The instructions do not currently enforce the use of strict delimiters or specific warnings to ignore embedded instructions when reading ingested files.
Capability inventory: The skill possesses the capability to execute maintenance scripts (Bash) and perform filesystem writes (Write/Edit).
Sanitization: While metadata fields like titles are sanitized for use in filenames, the content bodies are stored as raw markdown without filtering for agent-directed instructions.

llm-wiki