The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes several local and external command-line tools to perform its functions. It executes a bundled Node.js script (brand/materialize.mjs) for staging brand assets, calls the Higgsfield CLI for image generation, and uses npx playwright to render HTML templates into final PNG assets. These operations align with the skill's stated purpose.
[EXTERNAL_DOWNLOADS]: Branding metadata files (brand/brand.json and brand/finals.json) reference external assets such as fonts and logos hosted on Vercel's public blob storage. The caption template (brand/caption-skin.html) also loads the GSAP animation library from the JSDelivr CDN. These references target well-known, established services and are used solely for asset delivery.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates the creation of visual assets by interpolating user-provided text (such as headlines and descriptions) into an HTML template (references/overlay-shell.html). This template is subsequently rendered by the Playwright headless browser. This design presents a surface for indirect prompt injection if the processed data contains malicious code or instructions.
Ingestion points: User-provided inputs for {{KICKER}}, {{HEADLINE_HTML}}, {{CONTEXT}}, and {{ATTRIBUTION_OR_CONTEXT}} are inserted into the overlay HTML.
Boundary markers: The template does not employ explicit boundary markers or safety instructions to distinguish user data from the rendering logic.
Capability inventory: The skill can execute shell commands, perform image generation via Higgsfield, and render HTML via Playwright.
Sanitization: There is no evidence of HTML escaping or input validation performed on the user-supplied strings before they are rendered.

recoup-internal-consulting-generative-post