The Agent Skills Directory

[PROMPT_INJECTION] (HIGH): High-risk Indirect Prompt Injection surface. The skill defines a pipeline where untrusted data from the internet is intended to influence planning and execution logic.
Ingestion points: references/research-patterns.md and references/intelligence-rules.md explicitly define patterns for ingesting untrusted data via web_search and web_fetch tools.
Boundary markers: The skill employs extensive XML-based delimiting (e.g., <context>, <findings>, <research_scope>) which reduces accidental obedience but does not fully mitigate adversarial pressure.
Capability inventory: references/do-patterns.md contains templates that grant the agent the capability to create/modify files and execute shell commands (e.g., npm test, npx tsc, npx @redocly/cli).
Sanitization: No explicit sanitization or filtering of external content is defined within the templates.
[COMMAND_EXECUTION] (LOW): The skill templates promote the execution of subprocesses for verification. While these are standard development tools (npm, npx, tsc), they represent the 'execute' half of the dangerous read-then-execute pattern when paired with the research capabilities.

create-meta-prompts