The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill has a defined attack surface for indirect prompt injection as it ingests untrusted data from the web (Reddit, X, news sites, etc.). However, it mitigates these risks through structured processing and output sanitization.
Ingestion points: External data enters the agent context through search results returned by the nimble search tool in SKILL.md and references/sources.md.
Boundary markers: The skill instructions guide the agent to categorize and score data into discrete fields (reach, velocity, sentiment, risk) before rendering, which acts as a logical boundary.
Capability inventory: The skill environment allows for Bash execution, Write operations to the local ~/.nimble/ directory, and the spawning of sub-agents via the Agent tool.
Sanitization: The mandatory output templates (example.html and references/template.html) include explicit JavaScript functions (esc() and safeUrl()) to perform HTML entity escaping and protocol validation on all external content before rendering it in the user's browser.
[EXTERNAL_DOWNLOADS]: The skill references the vendor's own CLI tool and well-known libraries.
The installation guide in references/profile-and-onboarding.md points to the official @nimble-way/nimble-cli package on the NPM registry.
The HTML templates reference standard visualization libraries, such as D3 and TopoJSON, from trusted content delivery networks (cdnjs.cloudflare.com).
[DATA_EXfiltration]: The skill manages internal state and user profiles within the ~/.nimble/ directory. Network operations are limited to the Nimble search infrastructure and the retrieval of assets from well-known technology providers, with no evidence of sensitive system file access or credential harvesting.

brand-mention-monitor