The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests a user-provided 'concept prompt' (SKILL.md) and content from existing live websites (STEP 1) to generate code and infrastructure configurations. There are no explicit boundary markers or sanitization steps mentioned to prevent these inputs from influencing the agent's instructions during the site generation phase. Ingestion points: user-supplied and content fetched from the target . Boundary markers: none used to isolate untrusted input. Capability inventory: write access to the web monorepo, Git commit/push access, and infrastructure management via the Dokploy API. Sanitization: no explicit sanitization or validation of the generated content against the user's concept prompt.
[COMMAND_EXECUTION]: The skill relies on several scripts that execute system commands to process assets and manage the environment: 'scripts/dokploy.py' uses subprocess.check_output to execute the macOS security command to retrieve API keys; 'scripts/hero_video.mjs' uses execFileSync to run ffmpeg; and 'scripts/screenshot.sh' executes sips, ffmpeg, lsof, and launches a headless instance of Google Chrome.
[PRIVILEGE_ESCALATION]: The skill performs high-privilege operations, including modifying production Nginx configuration files, updating Dockerfiles, and pushing changes directly to a main Git branch. It also interacts with the Dokploy API to manage domains and environment variables. While these actions are central to the skill's purpose of shipping a website, they grant the agent significant control over the deployment environment.

one-prompt-website