The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from the public web via search-cli, WebSearch, and Exa MCP (Phase 3: RETRIEVE). It possesses high-impact capabilities including shell command execution and sub-agent spawning (Task tool), creating an attack surface for indirect prompt injection.
Ingestion points: Web results gathered through search tools and content extraction in reference/methodology.md.
Boundary markers: The skill contains instructions in reference/quality-gates.md to treat web content as data rather than instructions, but lacks strict programmatic isolation.
Capability inventory: Access to shell commands (bash, pnpm, python), file system writes (~/Documents/), and agent spawning via the Task tool.
Sanitization: Relies on LLM reasoning and qualitative checks; lacks rigorous input sanitization for gathered web text before synthesis.
[COMMAND_EXECUTION]: The skill workflow involves executing multiple local shell commands to facilitate the research process. SKILL.md instructs the agent to run pnpm run sync:research for publishing, and reference/methodology.md describes using search-cli and python scripts for validation (validate_report.py, verify_citations.py). Static analysis also identifies subprocess.run calls in the provided test suite (tests/test_citation_manager.py, tests/test_evidence_store.py).
[DATA_EXFILTRATION]: The skill automatically triggers a publishing process to lovstudio.ai after report generation. While intended as a feature, this involves transferring generated report data to a remote blog system via a local sync script.
[CREDENTIALS_UNSAFE]: README.md provides instructions for the user to configure sensitive API keys (e.g., Brave, Serper) for the search-cli tool. While it uses placeholders like YOUR_KEY, the design relies on the user managing these secrets in their environment for the tool to function.
[EXTERNAL_DOWNLOADS]: The skill encourages the installation of external dependencies, including search-cli from a third-party Homebrew tap (199-biotechnologies/tap) and weasyprint via pip. These tools are integrated into the core research and PDF generation workflows.
[PROMPT_INJECTION]: The SKILL.md file contains an 'Autonomy Principle' instructing the agent to 'Operate independently' and 'Only stop for critical errors', which encourages the agent to bypass human review steps for standard research tasks.

deep-research