The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes the claude CLI, a well-known tool from Anthropic, to perform automated code editing and testing. The execution is configured with specific restrictions, such as disabling persistence and blacklisting various Git commands (commit, push, merge, etc.) to ensure that the primary agent remains in control of the repository state.
[PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection because it reads and processes repository files that are then passed to an LLM-powered CLI tool with significant capabilities (e.g., Bash, file writing). This risk is addressed through the mandatory human-in-the-loop (primary agent) review process.
Ingestion points: Repository source files and git status information are read and included in implementation briefs passed to the claude CLI (SKILL.md).
Boundary markers: The skill employs a structured 'Implementation brief' template to provide clear instructions to the delegated tool.
Capability inventory: The delegated tool is granted Read, Edit, Write, Glob, Grep, and Bash capabilities via CLI flags (SKILL.md).
Sanitization: The workflow relies on post-execution verification, where the primary agent must independently inspect diffs and rerun validation suites before accepting changes.

claude-develop