The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines a mandatory "Agentic Protocol" (Step 2: Feynman-style Research) that requires the agent to use WebSearch tools to fetch real-world data and incorporate it into the reasoning process. This creates a surface where external, untrusted content from the web could potentially influence or override the agent's instructions.
Ingestion points: The agent is instructed to fetch technical principles, experimental data, and critical analysis from the web via the WebSearch tool (defined in SKILL.md).
Boundary markers: While the skill provides a strict "Answer Workflow" and "Fallback Tree," it lacks explicit delimiters or sanitization instructions for the external content fetched during the research phase.
Capability inventory: The agent has access to web search tools, local file writing, and shell execution (Git).
Sanitization: The instructions focus on accuracy ("don't fool yourself") but do not include security-focused sanitization of the input data.
[COMMAND_EXECUTION]: The "Version Self-Check" section in SKILL.md instructs the agent to execute shell commands to manage its lifecycle. Specifically, it directs the agent to run git rev-parse HEAD, git ls-remote origin HEAD, and suggest git pull --ff-only to the user. These commands are executed by the agent to compare local and remote versions of the skill.
[EXTERNAL_DOWNLOADS]: The README.md provides installation instructions that involve downloading content from external sources. These include using the npx skills add command (which utilizes a tool from the vercel-labs organization) and git clone targeting the author's own repository (alchaincyf/feynman-skill). These resources are hosted on GitHub, a well-known service, and originate from the skill's declared author.

feynman-perspective