The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill relies on dynamic execution to initialize its functionality within the agent's environment.
Evidence: The Setup section in SKILL.md instructs the agent to run exec(open("paper-narrative/kernel.py").read()) in a Python cell. This pattern executes the content of a local file directly in the active process context, which is a technique that can be used to bypass static analysis or hide complex behavior.
[INDIRECT_PROMPT_INJECTION]: The skill processes manuscript abstracts and figure metadata, which are untrusted external inputs, creating a surface for indirect prompt injection attacks.
Ingestion points: External content enters the skill via the abstract_text and figure_claims arguments in paper_brief_prompt and narrative_review_task within kernel.py.
Boundary markers: The untrusted data is separated from the instructions using only simple Markdown headers (e.g., ## Abstract, ## Figures), which can be easily bypassed by instructions embedded within the user data.
Capability inventory: The agent performs reasoning, role-playing (as a handling editor), and structured JSON generation based on the provided inputs.
Sanitization: There is no programmatic sanitization of the input text. The skill relies on a manual instruction to the agent to "re-read the whole brief" and verify the content as a mitigation strategy.

paper-narrative