The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Step 3 of the process in "references/01-process.md" instructs the agent to search the web and visit "discovered live references" to augment the local reference bank. This capability involves making network requests to arbitrary, unvetted domains.
[PROMPT_INJECTION]: The skill is susceptible to Indirect Prompt Injection. It ingests data from external websites to define brand attributes such as palette, typography, and voice (Step 4 of "references/01-process.md"). Malicious instructions on a visited site could manipulate the resulting brief or influence the agent's behavior.
Ingestion points: "references/01-process.md" Step 3 describes searching the live web and visiting external URLs.
Boundary markers: No instructions are provided to the agent to treat external content as untrusted or to ignore embedded instructions.
Capability inventory: The agent has file-writing capabilities and is instructed to commit discovered references back to its local knowledge base ("reference-bank/"), creating a vector for persistent injection.
Sanitization: No validation or sanitization is performed on the content retrieved from discovered URLs.
[DATA_EXFILTRATION]: The automated browsing of arbitrary URLs, driven by untrusted search results, creates a potential channel for the exfiltration of sensitive context if the agent's tool-use is compromised via malicious site content.

creative-brief-selector