The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The npx hyperframes capture <URL> command ingests content from external, untrusted websites to scaffold projects. This creates a surface where malicious instructions embedded in web content (e.g., hidden in HTML/assets or via vision-based captioning) could influence subsequent agent actions.
Ingestion points: references/init-and-scaffold.md via the capture command processing arbitrary URLs.
Boundary markers: The skill does not explicitly mention boundary markers for captured content, although BLOCKED.md is used to handle hard stops for failed captures.
Capability inventory: The skill possesses capabilities for file writing (init, capture), network operations (cloud render, publish, feedback), and shell command execution (npx commands).
Sanitization: Includes options to skip AI vision processing (--skip-vision) and uses cooperative budgets to limit post-navigation work.
[COMMAND_EXECUTION]: The skill makes extensive use of npx to execute the hyperframes CLI tool and its subcommands, which involves executing code downloaded from the npm registry.
[EXTERNAL_DOWNLOADS]: The browser ensure command downloads a pinned version of Chrome to ~/.hyperframes/. Additionally, deployment commands for AWS Lambda and Google Cloud Run interact with external infrastructure and package managers (e.g., bun, gcloud, terraform).
[CREDENTIALS_UNSAFE]: The skill manages and references sensitive credential files located at ~/.heygen/credentials and ~/.aws/credentials. While these are standard locations for CLI credentials, the skill actively interacts with these sensitive files to perform cloud rendering and deployments.
[DYNAMIC_EXECUTION]: The lambda deploy command dynamically builds a handler ZIP file at runtime using bun before deploying it to AWS infrastructure. The cloudrun command uses Terraform to generate and apply infrastructure configurations.

hyperframes-cli