The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches documentation and reference materials from Anthropic's official domain at https://code.claude.com/docs/en/. These are treated as well-known sources for the platform.
[REMOTE_CODE_EXECUTION]: The skill uses the uvx tool to download and execute the latest version of the skilllint package from a public registry during validation and debugging phases. This represents an unverifiable dependency execution at runtime.
[COMMAND_EXECUTION]: The skill runs a local Python script plugins/plugin-creator/scripts/create_plugin.py to generate plugin configuration files as part of the creation workflow.
[INDIRECT_PROMPT_INJECTION]: The skill has a significant exposure to indirect prompt injection due to its ingestion of untrusted external content that directly influences code generation and agent behavior.
Ingestion points: User-provided concepts in <plugin_target/>, captured preferences in discuss-CONTEXT.md, and consolidated research merged from external documents in research-FINDINGS.md.
Boundary markers: The skill uses XML-style tags to wrap some inputs, providing partial isolation within the prompt context.
Capability inventory: The skill is capable of writing new skill and agent files to the filesystem and executing shell commands via uvx and uv run.
Sanitization: No explicit sanitization or strict schema validation is described for the ingested content before it is processed by the agent.
[INDIRECT_PROMPT_INJECTION]: The skill explicitly instructs the AI agent to treat Mermaid process diagrams as "authoritative executable instructions" and an "executable instruction set". This elevates the risk of injection, as a maliciously crafted diagram in a researched file could override the agent's intended logic.

plugin-lifecycle