The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from live URLs, HTML/CSS code, and screenshots for usability auditing.
Ingestion points: Untrusted content enters the context via the /browse tool, the evidence-dir (reading page.html and screenshots), or directly provided code snippets.
Boundary markers: The skill contains robust instructions in references/orchestrated-runs.md and references/edge-cases.md explicitly telling the agent to ignore any instructions or commands embedded within the data being audited.
Capability inventory: The skill has the capability to write to the file system (Redesign mode), execute git commands (Repo Sync), and run a local Python analysis script.
Sanitization: All proposed code modifications are presented to the user as a dry-run diff, requiring explicit manual confirmation before any files are written.
[COMMAND_EXECUTION]: The skill executes shell commands for environment preflight checks, repository synchronization using git, and image processing via a local Python script (scripts/process_screenshots.py). These operations are scoped to the skill's primary purpose of UI auditing and are gated by user confirmation for any state-changing operations.
[EXTERNAL_DOWNLOADS]: The skill references and provides installation instructions for the agent-skill-manager (asm) and the browse tool from GitHub and NPM registries to facilitate dependency management.

dont-make-me-think