The Agent Skills Directory

[COMMAND_EXECUTION]: The utility script scripts/shots.mjs uses child_process.execSync to run a shell command (ls -d "$HOME"/.npm/_npx/*/node_modules/playwright-core) to locate the Playwright library within the user's local npx cache. While this command is used for environment discovery, it involves direct shell interaction with the user's home directory.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted external data, such as client design briefs, documents, and project source code. The scripts/design-audit.mjs tool generates prompts for sub-agents based on these assets, creating a surface where malicious content in a project file could influence agent behavior.
Ingestion points: SKILL.md (client briefs and assets), references/audit.md (project source code and route data).
Boundary markers: No explicit delimiters or instructions are provided to the LLM to ignore potentially malicious content embedded in the ingested assets.
Capability inventory: scripts/shots.mjs (automated browser navigation, shell command execution), scripts/design-audit.mjs (generates executable workflows and agent prompts).
Sanitization: The skill lacks explicit sanitization or filtering logic for external content before it is interpolated into audit prompts.
[DYNAMIC_EXECUTION]: The skill employs several dynamic execution patterns to facilitate its auditing features:
Dynamic Imports: scripts/shots.mjs uses dynamic import() to load playwright-core from filesystem paths computed at runtime via shell command output.
Browser Script Injection: scripts/shots.mjs uses ctx.addInitScript to inject key-value pairs provided by the user (via the --localstorage argument) directly into the automated browser context, which allows for arbitrary JavaScript execution within the test session.
Script Generation: scripts/design-audit.mjs programmatically generates a JavaScript workflow file (audit-workflow.js) which contains instructions and prompts for other agents.

prom-design