The Agent Skills Directory

[COMMAND_EXECUTION]: The skill dynamically generates multiple bash scripts (pack.sh, render-prompt.sh, copy-prompt.sh) and saves them to the .chalk/reviews/scripts/ directory. These scripts are then made executable using chmod +x and invoked via the Bash tool to perform git operations and manage prompt files. The reviewer argument is sanitized to a kebab-case string to prevent command injection at the prompt interface.
[PROMPT_INJECTION]: The skill possesses a surface for indirect prompt injection by incorporating external data into the generated AI prompts.
Ingestion points: The skill ingests data from the local environment via git diff, git log, and git status, as well as from local files such as handoff.md and user-supplied reviewer templates (SKILL.md).
Boundary markers: The generated prompt uses markdown headers (e.g., ## Review Pack, ## Handoff) and code blocks to separate different data sections. However, it lacks explicit instructions or system-level delimiters to prevent the reviewer agent from following instructions embedded within the codebase being reviewed.
Capability inventory: The skill utilizes the Bash, Write, Read, Glob, and Grep tools. It has the capability to write files, execute scripts, and interact with the system clipboard using tools like pbcopy, xclip, or wl-copy (SKILL.md).
Sanitization: The reviewer name is sanitized to prevent shell injection, but the contents of git diffs and handoff files are interpolated directly into the final prompt file without further sanitization or escaping of potentially malicious instructional content.

create-review