The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill recommends installing '@anthropic-ai/claude-code' from the official NPM registry, which is a package maintained by a trusted organization.
[COMMAND_EXECUTION]: The skill constructs and executes shell commands for 'git' operations and the 'claude' CLI. It uses double-quoting for shell variables like '$PROMPT' and '$DIFF' in SKILL.md to mitigate direct command injection.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it processes untrusted external data. 1. Ingestion points: Git diffs, log statistics, and commit history are ingested into the agent context via shell variables in SKILL.md. 2. Boundary markers: The prompt templates in SKILL.md and adversarial-prompt.txt use textual labels like 'Diff:' and 'REPOSITORY CONTEXT:' to separate the diff from instructions, but these are not robust against adversarial content. 3. Capability inventory: The sub-instance of Claude is explicitly granted 'Read' and 'Bash' tool permissions and is configured with '--dangerously-skip-permissions' to allow autonomous execution. 4. Sanitization: There is no evidence of validation or sanitization for the code diffs before they are interpolated into the prompt.

claude-cli