The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains instructions to bypass user confirmation, specifically stating "No asking. Stage everything and commit immediately." This overrides the agent's standard behavior of seeking user approval before performing write operations.
[COMMAND_EXECUTION]: The skill utilizes dynamic context injection in SKILL.md to execute shell commands (git branch, git status, git diff HEAD, and git log) at the moment the skill is loaded. This allows the skill to read repository data into the LLM context without a specific user trigger.
[COMMAND_EXECUTION]: The skill uses git add -A, which stages all changes in the working directory. This increases the risk of accidentally committing sensitive files, such as .env files or hardcoded credentials, especially since it instructs the agent not to ask for confirmation.
[PROMPT_INJECTION]: The skill is vulnerable to Indirect Prompt Injection because it processes untrusted data from the local repository (file diffs and commit logs) without sanitization or boundary markers.
Ingestion points: Untrusted data enters the context via !git diff HEAD and !git log in SKILL.md.
Boundary markers: Absent. The output of the git commands is interpolated directly into the context without delimiters or instructions to ignore embedded commands.
Capability inventory: The skill has the capability to execute git add -A and git commit via shell commands defined in the Steps section.
Sanitization: Absent. There is no filtering or validation of the content retrieved from the git repository before it is processed by the agent.

lgnh-commit-detail