The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external, potentially untrusted text to extract entities, relations, and events, creating a surface for indirect prompt injection.
Ingestion points: According to references/extraction.md, the skill ingests "unstructured text, transcripts, and PDFs" from users to build knowledge graphs.
Boundary markers: The skill uses markdown tags like <text> and <ontology> in its prompt patterns to delimit untrusted content, which provides a basic layer of protection but is not a robust defense against adversarial payloads.
Capability inventory: The agent has the capability to generate visual artifacts (Mermaid diagrams and HTML pages) and orchestrate subagents through "teammode" or independent parallel nodes, which could be manipulated by malicious instructions in the input text.
Sanitization: There is no evidence of input sanitization or filtering to prevent the agent from interpreting embedded instructions within the processed datasets.
[EXTERNAL_DOWNLOADS]: The skill's task orchestration logic depends on an unverified external source for its runtime implementation.
Evidence: In references/task-graphs.md, the skill references the lazycodex repository (github.com/code-yeongyu/lazycodex) as the "runtime basis" for its teammode member inputs and wave mapping.
Concern: Directing an agent to follow logic or tool definitions hosted in an untrusted third-party repository introduces a supply chain risk, as the external code could be modified to include malicious behaviors without the user's knowledge.

graph-engineering