The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill handles external data in the form of academic paper titles, abstracts, and full texts retrieved during literature discovery steps. This creates an implicit attack surface where malicious or adversarial text within an academic paper's abstract or title could influence the subsequent multi-LLM generation pipeline.
Ingestion points: Untrusted inputs enter the context via variables like [PAPER LIST], [OUTLINE LIST], and [SUBSECTION] across various prompts in references/survey-prompts.md.
Boundary markers: The prompt templates employ structured blocks with --- dividers and XML-style tags (such as <instruction> and <format>) to isolate untrusted text from instructions.
Capability inventory: The workflow relies on executing a local python script ~/.claude/skills/deep-research/scripts/search_semantic_scholar.py to collect research papers, but does not perform administrative operations or arbitrary code generation.
Sanitization: Content fields are interpolated directly into prompts without specific sanitization mechanisms or filtering filters for prompt injection patterns.

survey-generation