The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes project-level files (PRODUCT.md, DESIGN.md) and browser-supplied data (manual edits) into the agent's context. This creates an attack surface where instructions embedded in project documentation or browser events could influence agent behavior.
Ingestion points: PRODUCT.md, DESIGN.md (via scripts/context.mjs), and browser events/manual edits (via scripts/live-server.mjs).
Boundary markers: Absent in project documentation files; the agent relies on its own instruction boundaries.
Capability inventory: File system modification (fs.writeFileSync), arbitrary command execution (execFileSync, spawn), and network requests (fetch).
Sanitization: Minimal sanitization of external text before it is interpolated into prompts for implementation or refinement.
[COMMAND_EXECUTION]: The skill frequently executes local binaries and spawns child processes to facilitate its design and audit features.
Evidence: scripts/live-copy-edit-agent.mjs spawns AI CLIs (Claude and Codex) with flags like --dangerously-bypass-approvals-and-sandbox to apply code changes. scripts/live-server.mjs and other scripts execute node, git, and python to inspect the repository and manage the live session.
Context: These high-privilege operations are associated with the skill's primary purpose of automating frontend design tasks.
[EXTERNAL_DOWNLOADS]: The skill performs network operations to check for updates.
Evidence: scripts/context.mjs performs a fetch request to https://impeccable.style/api/version to determine if a newer version of the skill is available. This is a non-whitelisted domain belonging to the tool's vendor.

impeccable