The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data (manuscripts, abstracts, and captions) and interpolates them directly into LLM prompts within its JavaScript control-plane logic.
Ingestion points: The variables manuscriptText, abstractText, and captionsText are loaded from Artifact Versions and concatenated into the prompt string in SKILL.md (Step 1).
Boundary markers: The prompt construction lacks explicit delimiters (such as XML-style tags or structured barriers) to separate the system instructions from the potentially adversarial content within the manuscript text.
Capability inventory: The skill possesses the ability to execute code via notebook_execute, delegate tasks to other skills using host.delegate, and write new project artifacts.
Sanitization: No sanitization, filtering, or instruction-override warnings are applied to the input text before it is sent to the model.
[COMMAND_EXECUTION]: The skill utilizes platform-specific execution capabilities to perform its tasks.
SKILL.md documents the use of notebook_execute to call deterministic Python functions defined in kernel.py (e.g., paper_brief_schema).
The JavaScript logic uses host.delegate and host.collect to manage child agent workflows, which is an expected but powerful capability for workflow orchestration.

paper-narrative