The Agent Skills Directory

[SAFE]: The skill is architected as a read-only curation layer that identifies library issues and emits preview plans. It explicitly prohibits the agent from performing CRUD (Create, Read, Update, Delete) operations, ensuring that no data modifications occur without manual user intervention or specific external command invocation.- [COMMAND_EXECUTION]: The instructions direct the agent to perform basic environmental checks using 'ls' and 'research-hub doctor' to verify the presence of necessary sibling skills or CLI tools before attempting to audit the library. This is a standard and safe diagnostic practice.- [EXTERNAL_DOWNLOADS]: The skill references installation commands for 'research-hub-pipeline' and the 'zotero-skills' repository. These are author-owned tools required for Zotero connectivity and are documented for user-led environment setup.- [DATA_EXFILTRATION]: The skill reads Zotero library metadata via a local API and accesses internal cluster configuration files (e.g., '.research_hub/clusters.yaml'). This data access is restricted to the local environment and is strictly required for the skill's stated purpose of auditing the bibliography library.- [PROMPT_INJECTION]: An indirect prompt injection surface exists because the skill processes untrusted bibliographic data (paper titles, tags) from Zotero. However, the risk is negligible as the skill lacks automated write capabilities and is explicitly constrained to outputting formatted reports for human review. Mandatory evidence chain for Category 8: 1. Ingestion points: Zotero local API, .research_hub/clusters.yaml, .research_hub/dedup_index.json. 2. Boundary markers: Absent in instructions, but enforced via output report templates. 3. Capability inventory: Local file reads and Zotero API reads. 4. Sanitization: Absent, relies on LLM formatting and human review of the generated preview plan.

zotero-library-curator