The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill interacts with untrusted external content from websites and native applications, which could contain malicious instructions.
Ingestion points: Web page content (via Playwright) and application interfaces (via computer-use) as defined in SKILL.md.
Boundary markers: The skill includes explicit instructions in the 'Prompt injection boundary' sections to disregard instructions found within the tested content.
Capability inventory: The skill has access to shell execution (bash), browser automation (Playwright), and system interaction (computer-use).
Sanitization: Instructions require the agent to log suspicious content as a finding rather than following it.
[COMMAND_EXECUTION]: The skill invokes several shell scripts located in ~/.claude/skills/nanostack/bin/ for telemetry, session management, and artifact storage.
[DYNAMIC_EXECUTION]: The bin/screenshot.sh script uses node -e to execute a JavaScript string for browser automation, which is used to capture test evidence.
[EXTERNAL_DOWNLOADS]: The skill relies on the playwright Node.js package, a well-known tool for browser automation, which downloads browser binaries during setup.

qa