The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted external inputs, specifically user-supplied manuscript drafts and unstructured journal reviewer feedback, as documented in agents/pipeline_orchestrator_agent.md and references/external_review_protocol.md. There are no explicit boundaries, delimiters, or guidelines instructing the processing models to treat the embedded content strictly as passive data. The skill's capability inventory includes the execution of shell utilities via the Bash tool, alongside file modification capabilities (Write, Edit) and network access (WebSearch, WebFetch). No validation or input escaping is implemented to sanitize data before passing it to sub-agents or scripts.
[COMMAND_EXECUTION]: The orchestration workflow in agents/integrity_verification_agent.md executes local Python scripts (verify_citations.py and claim_faithfulness.py) via the Bash tool to handle automated reference and claim validation.

alterlab-research-pipeline