The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill documentation and its bootstrap scripts encourage the installation of CLI tools via curl | bash pipes from remote servers.
Evidence in SKILL.md: curl -fsSL https://raw.githubusercontent.com/yschimke/skills/main/scripts/install.sh | bash.
Evidence in scripts/compose-preview: curl -fsSL "https://raw.githubusercontent.com/yschimke/skills/main/scripts/install.sh" | bash -s -- --cli-only.
Evidence in references/agent-cloud.md: curl -fsSL https://env.coo.ee/compose | bash and curl -fsSL https://raw.githubusercontent.com/yschimke/skills/main/scripts/install.sh | bash -s -- --android-sdk.
Evidence in references/cli.md: curl -fsSL https://bb.staticvar.dev/install.sh | bash.
[INDIRECT_PROMPT_INJECTION]: The skill reads local project source files to discover previews and renders UI components, which could contain malicious prompts meant to manipulate agent behavior.
Ingestion points: Tools such as find_previews_for_file read local Kotlin source files and the render_preview tool processes UI content.
Boundary markers: No specific delimiters or instructions to ignore embedded prompts in the analyzed data were identified in the agent instructions.
Capability inventory: The skill allows the agent to execute shell commands, run Gradle tasks, and perform network operations via the catalog MCP.
Sanitization: No evidence of sanitization or filtering of processed source code or rendered UI elements was found.
[COMMAND_EXECUTION]: The skill makes extensive use of local shell commands to manage the Gradle build system and drive the rendering CLI tool.
[DYNAMIC_EXECUTION]: The skill employs a Gradle plugin that scans and executes code from compiled classes. It also uses 'Remote Compose' to compile and replay UI trees from document byte streams.

compose-preview