The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes an instruction to "run silently, don't ask" when executing a skill update command (npx hyperframes skills update talking-head-recut). This is a direct attempt to override the agent's safety protocols and bypass user oversight for shell command execution.
[REMOTE_CODE_EXECUTION]: The skill uses npx to fetch and execute the hyperframes tool from a remote registry. The instruction to automate updates without user approval increases the risk that malicious code could be downloaded and executed in the host environment.
[COMMAND_EXECUTION]: The skill makes extensive use of powerful shell utilities, including ffmpeg, ffprobe, and several npx subcommands for transcription and rendering. The request for these to be executed autonomously poses a significant security concern.
[DYNAMIC_EXECUTION]: The agent is instructed to write HTML and CSS fragments that are subsequently interpreted and rendered. While the skill defines rules to exclude <script> tags, the generation of executable layout content provides a surface for potential exploitation.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection due to its processing of external video transcripts and user briefs.
Ingestion points: The skill reads transcript.json (generated from source audio) and BRIEF.md (user-provided instructions).
Boundary markers: There are no explicit delimiters or instructions provided to the agent to ignore potentially malicious content within these data sources.
Capability inventory: The skill has access to shell execution (npx, ffmpeg), file writing (public/cards/*.html), and network operations (via npx updates).
Sanitization: The skill references a "HyperFrames sanitizer" for CSS and manually instructs the agent to avoid <script> tags and external URLs, but lacks robust programmatic enforcement of these boundaries.

talking-head-recut