The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external websites and local project files to create a video storyboard and plan. This creates a significant surface for indirect prompt injection attacks.
Ingestion points: The skill specifically instructs the agent to read external content from http(s):// URLs and local source code (main page, routes, README).
Boundary markers: The instructions do not provide delimiters or guidance to the agent to ignore instructions embedded within the site's text or code comments.
Capability inventory: The skill empowers the agent to write and execute scripts, perform network operations to download assets, and write files to the local directory.
Sanitization: There is no mention of sanitizing or escaping the captured text before it is used in the brag-plan.md or interpolated into scripts.
[DYNAMIC_EXECUTION]: The skill requires the agent to generate and execute code at runtime to produce the video deliverables.
Evidence: The instructions state: "You make the whole video yourself... with whatever tools are on the machine" and "Keep every intermediate file (frames, downloads, scripts, stems) in a work/ subfolder". This encourages the creation of temporary scripts that are executed to perform rendering and assembly.
[COMMAND_EXECUTION]: The skill relies on the agent invoking local CLI tools for media processing.
Evidence: Instructions such as "Build it with whatever works on this machine" and "render brag.mp4" imply the use of subprocesses (e.g., ffmpeg, node, python) to handle file manipulation and video encoding.
[EXTERNAL_DOWNLOADS]: The skill performs network operations to fetch assets from arbitrary remote servers.
Evidence: Under the Website section, the skill directs the agent to download the logo, product screenshots, hero images, and demo videos into the work/ directory.

brag-slim