The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content including transcripts, OCR, and metadata. It proactively mitigates injection risks by defining an 'untrusted-content-contract' that forbids external data from authorizing commands or being interpolated into executable contexts.
Ingestion points: transcripts/{platform}/{id}.txt, frame-analysis/{platform}/{id}.json, and metadata.json.
Boundary markers: Explicit instructions to treat all external strings as untrusted and keep them delimited during classification.
Capability inventory: File writes (analysis/*.json), package installation (npm install), and local web server execution (python -m http.server).
Sanitization: Mandatory use of textContent and document.createElement() for DOM updates; strict avoidance of innerHTML and eval equivalent sinks.
[EXTERNAL_DOWNLOADS]: The skill downloads the chart.js package using npm.
Evidence: Uses npm install --ignore-scripts --save-exact chart.js@4.5.1. The use of --ignore-scripts is a security best practice to prevent execution of malicious lifecycle scripts during installation.
[COMMAND_EXECUTION]: The skill uses standard development commands for dependency management and testing.
Evidence: npm install, mkdir, cp, and python -m http.server. All commands are restricted to the local environment and do not involve piping remote content to a shell.
[CREDENTIALS_UNSAFE]: No hardcoded credentials or unsafe secret handling practices were detected. The skill specifically instructs that external content cannot authorize the use of credentials.

video-dashboard