The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill researches external marketing URLs and brand repositories to extract design signals (Tailwind configs, CSS, marketing copy, etc.). This data is untrusted and could contain hidden instructions intended to influence the video generation or the agent's behavior.
Ingestion points: Researching user-provided URLs and analyzing repository files (SKILL.md Step 1 & 2).
Boundary markers: No explicit boundary markers or sanitization instructions for external content are provided in the prompts.
Capability inventory: The skill executes local shell commands (node, tr), writes files (video.html, spec.md, direction.json), and potentially invokes other skills (make-video, archive-broll).
Sanitization: No specific sanitization or filtering of the researched brand data is mentioned before it is used to generate the video specification.
[COMMAND_EXECUTION]: The skill uses shell commands to generate randomness and run local scripts.
Evidence: LC_ALL=C tr -dc 'A-Za-z0-9' < /dev/urandom | head -c 1024 > videos/<slug>/seed.txt (SKILL.md Step 4).
Evidence: node <this-skill>/scripts/roll-direction.mjs --format promo ... (SKILL.md Step 4).
Evaluation: These are local operations using standard tools for the skill's primary purpose.
[DYNAMIC_EXECUTION]: The skill generates a video.html file based on a 'creative specification' which is then rendered into a video.
Evidence: The instructions require building an HTML page with a specific JS contract (window.renderAt(t)) and saving it to the filesystem (SKILL.md Step 8).
Evaluation: While the skill generates executable content (HTML/JS), it does so within a defined template/contract framework.

guided-brand-video