The Agent Skills Directory

[Command Execution]: The skill executes various system commands to initialize the project environment, including npm install for dependency management, git init for version control, and node to run internal helper scripts for site provisioning (documented in SKILL.md, Phase 2 and Phase 1.5). These commands are necessary for the skill's primary function but involve the execution of external binaries.
[External Downloads]: The skill fetches assets and scripts from external sources to support site functionality and quality: it retrieves high-quality photos from Unsplash (Phase 5.3) and downloads the axe-core accessibility testing library from a well-known content delivery network (scripts/axe-audit.js). These are established services used to fulfill the skill's stated purpose.
[Remote Code Execution]: As part of its quality assurance process, the skill injects and runs the axe-core script within a browser environment using Playwright (evidence in scripts/axe-audit.js: await page.addScriptTag({ url: AXE_CDN_URL });). This is a standard procedure for automated accessibility auditing.
[Indirect Prompt Injection]: The skill accepts user input for site metadata and substitutes it into the project's source code, which could potentially be used as an injection vector for malicious instructions:
Ingestion points: Site purpose, name, and description are gathered from $ARGUMENTS and AskUserQuestion prompts in Phase 1 (SKILL.md).
Boundary markers: The skill uses approval gates in Phase 4 and Phase 7 to allow the user to review the generated content, but does not employ specific delimiters to isolate interpolated user data from the rest of the prompt context.
Capability inventory: The skill possesses extensive capabilities including file writing (Write, Edit tools) and command execution (Bash tool) across all implementation phases (SKILL.md).
Sanitization: While the skill performs validation to ensure placeholders are replaced (scripts/validate-site.js), it does not explicitly describe automated sanitization of user-provided strings against prompt injection techniques.

create-site