The Agent Skills Directory

[COMMAND_EXECUTION]: The skill scaffolds a .claude/settings.local.json file which grants pre-approved execution permissions for several CLI tools including ddev, git, and gh (GitHub CLI). This enables the agent to perform project management tasks, such as commits and pull requests, without requiring manual user approval for every command.
[EXTERNAL_DOWNLOADS]: The skill facilitates the installation of the @anthropic-ai/chrome-devtools-mcp tool via npx and uses git clone to fetch public repositories into an ephemeral research/ directory for code audits. These operations target well-known services and trusted repositories.
[PROMPT_INJECTION]: The skill presents an indirect prompt injection surface as it ingests data from untrusted local files like composer.json and package.json to customize its output.
Ingestion points: Project detection phase reads composer.json, package.json, .ddev/config.yaml, and git logs.
Boundary markers: Absent; the skill directly uses detected values for template interpolation.
Capability inventory: The skill can write configuration files and grant broad execution permissions via settings.local.json.
Sanitization: No explicit sanitization or validation of the metadata read from project files before it is placed into generated rules.
[COMMAND_EXECUTION]: The generated rule templates explicitly instruct the agent to use absolute paths for git commands. This is a security best practice intended to prevent the accidental execution of untrusted git hooks that might be present in a repository's local environment.

craft-project-setup