The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill documentation instructs the agent to install numerous third-party plugins and packages from non-trusted repositories. Specifically, it references github.com/cursor/plugins and github.com/PriNova/pi-agent-codebase-workflows for critical workflow stages.
[REMOTE_CODE_EXECUTION]: Several files describe remote installation and execution patterns that bypass traditional package management. For example, references/cli-tools/agent_browser.md suggests using npx -y @earendil-works/pi-agent-browser, which downloads and executes code from NPM on demand. Other tools are installed via pi install git:... which pulls directly from GitHub.
[PROMPT_INJECTION]: The 'Weekly Intelligence Canvas' variant in SKILL.md constitutes an indirect prompt injection surface (Category 8). It explicitly instructs the agent to ingest and analyze data from untrusted external sources including company blogs, GitHub changelogs, LinkedIn, X (Twitter), and Reddit.
Ingestion points: SKILL.md Variant 2 (Search across blogs, social media, news, and forums).
Boundary markers: Absent from the prompt template provided for Variant 2.
Capability inventory: The agent has extensive capabilities across the skill files, including bash command execution, write access, subagent delegation, and web browsing via agent_browser.
Sanitization: No sanitization or filtering logic is specified for the ingested web content.
[COMMAND_EXECUTION]: The skill uses several potentially dangerous command-line patterns for system operations. For instance, references/cli-tools/read-config.md uses node -e to dynamically execute code for parsing configuration files, and references/cli-tools/file-locking.md uses shell scripts with ln, sha1sum, and cat for managing file locks.

stelow-product-multi-method-market-analysis