The Agent Skills Directory

[PROMPT_INJECTION]: The skill instructions in reference/overdrive.md and reference/bolder.md contain strong directives to the AI agent to ignore its default behaviors or training reflexes (e.g., 'Reject them first', 'forbid the trio from sharing theme'). While these are aimed at improving design quality and avoiding 'AI slop', they represent instructions intended to override standard model responses.
[COMMAND_EXECUTION]: The skill utilizes a local HTTP server (scripts/live-server.mjs) and several helper scripts to facilitate live iteration. These scripts execute shell commands, spawn background processes, and perform file system operations to inject code into the user's project.
[EXTERNAL_DOWNLOADS]: The skill uses npx impeccable to execute detection and cleanup tasks. This utility downloads and executes packages from the npm registry at runtime. It also references external design assets from the author's GitHub repository and stock imagery from Unsplash.
[INDIRECT_PROMPT_INJECTION]: The skill has a significant ingestion surface for indirect instructions via project files and browser data.
Ingestion points: The skill reads PRODUCT.md, DESIGN.md, and project source files (HTML, CSS, JS) via scripts/load-context.mjs and reference/critique.md.
Boundary markers: Absent. The agent is not instructed to use delimiters or ignore instructions within the ingested project files.
Capability inventory: The skill can execute arbitrary shell commands via the Bash tool, write or edit project files, and maintain a local network server.
Sanitization: The skill performs minimal sanitization of ingested content, such as the cssSafe function in scripts/design-parser.mjs for design token values.
[PERSISTENCE_MECHANISMS]: The scripts/pin.mjs utility allows the skill to create persistent standalone shortcuts within the AI agent's harness directories (e.g., .claude/skills/). This modifies the agent's environment to add new invocable commands.

iblai-vibe-design