oil-oil/kimi ? reverse-engineered prompt

Build me a small agent skill that makes it easy for any coding agent to delegate focused tasks to a local Kimi Code CLI from the terminal.

I want it to work on macOS and Linux with Bash, use a simple shell wrapper, and let the caller pass a task plus a workspace and a few important files or images for context. It should run Kimi in a non interactive way, capture progress in a clean way, save the final answer as a compact Markdown file, and return a session id so the same conversation can be resumed later. The main goal is reliable delegation for UI, layout, design system, and frontend polish work, not just generic coding help.

Please include clear setup instructions, a skill file for agents to load, and a script that is easy to validate locally. If you need to check current Kimi CLI behavior or flags, look up the latest docs online first.