The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions encourage the use of CLI flags that disable interactive security prompts for file edits and shell commands.
Evidence in SKILL.md: Instructions recommend using --permission-mode bypassPermissions for coding tasks and document --dangerously-skip-permissions as a way to "directly skip all permission checks."
These configurations allow the agent to perform potentially destructive actions, such as modifying source code or executing system commands, without explicit user confirmation during the automated session.
[EXTERNAL_DOWNLOADS]: Fetches and installs the official Claude Code CLI from the developer's public package registry.
Evidence in SKILL.md: npm install -g @anthropic-ai/claude-code.
This is a standard prerequisite for the skill's operation and targets a well-known service provider.
[PROMPT_INJECTION]: The skill implements a delegation pattern that processes arbitrary task descriptions in a secondary agent session, creating a surface for indirect prompt injection.
Ingestion points: The task description passed to the claude -p command in SKILL.md.
Boundary markers: The skill uses XML-style tags in its prompt recipes (e.g., in references/task-prompt-recipes.md) to structure instructions, but does not include explicit delimiters or warnings to ignore instructions embedded in the processed data.
Capability inventory: The delegated session has access to the full suite of Claude Code tools, including file reading/writing and bash execution.
Sanitization: No sanitization or validation of the task input is performed before it is passed to the sub-session.

claude-code-agent