The Agent Skills Directory

[PROMPT_INJECTION]: The skill ingests untrusted data from web pages (DOM snapshots and console logs) which could contain malicious instructions meant to influence agent behavior.
Ingestion points: The tools extension_source_inspect and extension_logs read data directly from the browser context (documented in references/debugging.md).
Boundary markers: The documentation suggests marking injected roots with [data-extension-root] to distinguish extension content from page content.
Capability inventory: The skill has access to tools for file creation (extension_create), code evaluation (extension_eval), and network publishing (extension_publish).
Sanitization: Core Rule 17 in SKILL.md explicitly instructs the agent to treat page content as untrusted input and recommends using the --source-redact flag to mitigate risk.
[COMMAND_EXECUTION]: The skill uses the extension CLI and MCP server to manage the development lifecycle, including scaffolding and building projects.
Evidence: Uses npx extension@latest create for project setup and npm run build for artifact generation.
[EXTERNAL_DOWNLOADS]: The skill references and downloads project templates from well-known public repositories.
Evidence: Fetches project templates from the official extension-js/examples GitHub repository (documented in references/templates.md).

extension-dev