The Agent Skills Directory

[COMMAND_EXECUTION]: The skill attempts to locate and execute a local binary named 'browse' from multiple paths, including within the agent's skill directory and the user's home directory (~/.claude/skills/...). It uses this binary to perform web snapshots and screenshots.
[DATA_EXFILTRATION]: The skill reads from '~/.lazyweb/libraries.json', which may contain configuration or session data for third-party design platforms. While used for functional library search, this involves accessing local configuration outside the skill's immediate directory. Additionally, screenshots of the user's design are converted to Base64 and sent to an external MCP tool ('lazyweb_compare_image') for visual search.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted content from competitor websites and external design libraries to generate design critiques and improvement ideas. An attacker could place malicious instructions on a website that the agent researches to influence the report output or agent behavior.
Ingestion points: Competitor websites identified via search, connected design libraries (e.g., Mobbin, Savee), and Lazyweb database results (SKILL.md).
Boundary markers: None identified; the agent is instructed to read 'visionDescription' and competitor content to generate ideas.
Capability inventory: Execution of local binaries (Bash), network requests (curl), and file creation (Write).
Sanitization: No evidence of sanitization or safety filters for processed web content.
[EXTERNAL_DOWNLOADS]: The skill uses 'curl' to download reference images from external URLs ('imageUrl') and save them to the local filesystem.

lazyweb-design-improve