The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a process where user-supplied answers from an interview (Step 0) are written verbatim into a BRIEF.md file. This content is later processed by the agent using vision_analyze and browser-based verification tools, creating an attack surface for indirect prompt injection if the user provides malicious instructions disguised as brand information.
Ingestion points: SKILL.md (Step 0) and resulting BRIEF.md in the workspace.
Boundary markers: None present; content is written "verbatim".
Capability inventory: write_file, vision_analyze (on screenshots), and browser execution via Playwright.
Sanitization: No sanitization or escaping is performed on the user-supplied strings before they are incorporated into the build artifacts.
[COMMAND_EXECUTION]: The skill relies on executing system commands and scripts to perform its core functions.
Evidence: scripts/doctor.mjs and scripts/encode.sh invoke ffmpeg and ffprobe to validate the environment and encode video assets.
Evidence: SKILL.md instructs the agent to run Node.js scripts (serve.mjs, shoot.mjs) which use subprocesses and browser automation.
[EXTERNAL_DOWNLOADS]: The skill communicates with external, non-whitelisted APIs to generate AI assets.
Evidence: scripts/kie.mjs makes network requests to api.kie.ai and kieai.redpandaai.co to upload images and generate video content using the KIE_AI_API_KEY provided in the environment.
[DYNAMIC_EXECUTION]: The skill uses dynamic loading and runtime code evaluation to facilitate its verification harness.
Evidence: scripts/shoot.mjs and scripts/doctor.mjs use createRequire with computed paths to resolve playwright-core from the project's local node_modules directory.
Evidence: scripts/shoot.mjs uses Playwright's page.evaluate() to execute JavaScript within the context of the generated landing page to extract state and perform contrast checks.

scrollcraft