The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The SKILL.md file includes a prerequisite check that references a shell-piped installation command (curl -LsSf https://astral.sh/uv/install.sh | sh) for the uv tool. This command downloads and executes a setup script from Astral's official domain. While Astral is the legitimate vendor for this well-known tool, the use of piped execution for remote scripts is a high-privilege operation.
[INDIRECT_PROMPT_INJECTION]: The skill's workflow involves a 'Research pass' where it uses agent-browser to visit and snapshot web pages. The data from these snapshots is used by the agent to construct a script.json file that dictates subsequent browser actions and narration. Malicious content on a researched web page could potentially influence the agent's generation of this script, leading to unintended browser interactions.
Ingestion points: Accessibility tree snapshots and page metadata retrieved via agent-browser snapshot -i during the research phase.
Boundary markers: None. There are no explicit instructions to ignore embedded commands or delimiters used when the agent processes the researched content.
Capability inventory: Full browser automation (navigation, clicking, form filling) and arbitrary JavaScript execution via eval within agent-screencast.py; command-line execution of ffmpeg for video assembly.
Sanitization: None. The skill relies on the agent to interpret the researched content safely without a dedicated validation or filtering step.
[COMMAND_EXECUTION]: The agent-screencast.py script executes external binaries including agent-browser, ffmpeg, and ffprobe using subprocess.run. Although the script correctly passes arguments as lists and uses shlex.split() to prevent simple shell injection, the content of these arguments is derived from the agent-generated script.json, which is a downstream product of untrusted web research.
[DYNAMIC_EXECUTION]: The skill implements an eval command that allows the execution of arbitrary JavaScript within the automated browser session. The agent-screencast.py script facilitates this by piping strings from the script.json into the agent-browser tool's stdin. This provides a direct path for dynamic execution of logic generated by the agent based on its environment.
[EXTERNAL_DOWNLOADS]: The skill depends on multiple external tools and libraries, including the agent-browser NPM package and the uv Python tool. The Python script also specifies several third-party dependencies (edge-tts, google-genai) in its inline metadata that are retrieved from public package registries.

agent-screencast