The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The deck.py script connects to api.cloudflare.com and api.typesafe.ai to fetch scoring data used to rank and select 'lenses' for the agent council.
[DATA_EXFILTRATION]: The skill sends the text of the user's question or problem description to external APIs (api.cloudflare.com and api.typesafe.ai). While Cloudflare is a well-known service, the transmission of arbitrary user input to external endpoints is a data exposure risk.
[INDIRECT_PROMPT_INJECTION]: The skill is a multi-agent orchestrator that ingests untrusted user input (the problem/question) and interpolates it directly into the prompts for sub-agents (e.g., in brainstorm-agent-prompt.md, perspectives-prompt.md, etc.).
Ingestion points: User-supplied --question or PROBLEM strings enter the context via deck.py and markdown templates.
Boundary markers: The templates use formal labels (e.g., PROBLEM:, CONSTRAINTS:) but lack robust 'ignore embedded instructions' delimiters.
Capability inventory: Sub-agents are granted access to tools like web search, file reading, and shell execution (in references/harness.md).
Sanitization: No explicit sanitization or escaping of the user-provided question text was observed before interpolation.
[COMMAND_EXECUTION]: The deck.py script executes local commands such as cf auth whoami to manage Cloudflare authentication for its scoring features.

heavy-think