The Agent Skills Directory

[SAFE]: The skill implements a research-oriented workflow that manages local research artifacts and academic theory search. It relies on explicit user confirmation at multiple critical gates (Source Problem Gate, Problem-Theory Decomposition, Theoretical Tradition Selection) to ensure that all conclusions and framing changes are reviewed by the researcher. This human-in-the-loop design effectively manages the risk of automated errors.
[PROMPT_INJECTION]: The skill processes potentially untrusted content from research papers and search results (SKILL.md Step 8, 12). This represents an indirect prompt injection surface. The evidence chain for this finding is as follows: 1) Ingestion points: Local research files and external theory sources are read into the agent context in Steps 8 and 12. 2) Boundary markers: The workflow utilizes markdown templates to structure output, though no explicit delimiters or safety warnings for untrusted content are present. 3) Capability inventory: The skill is authorized to perform file-write operations within the project workspace; no arbitrary code execution or exfiltration capabilities were identified. 4) Sanitization: No content validation or sanitization of ingested research data is performed. This risk is inherent to the primary research task and is mitigated by the structured workflow.

paper-reading-theoretical-grounding