boringmarketer/kimi-first

kimi-first is a skill for Claude Code, the AI coding assistant, that teaches it to hand off certain implementation work to a second AI tool called the Kimi Code CLI. The idea is a division of labor: Claude keeps doing the judgment heavy parts of a coding task, like planning specs, designing architecture, coordinating work, reviewing results, and pressing the final merge button, while Kimi does the hands on typing for tasks where the plan is already frozen and there are no more decisions left to make, such as building out an agreed spec, doing repetitive migrations, filling in tests, sweeping many files with the same change, or fixing failing checks. The project is a direct port of another skill called codex-first, originally built for a different CLI tool named Codex, with the same underlying approach just swapped over to work with Kimi instead. When Claude decides a task is ready to be delegated, it writes out a detailed prompt to a temporary file covering the goal, the relevant file paths, any constraints, what should not be touched, and what proof of success looks like. It then runs the Kimi tool in headless mode inside the target project, with Kimi's tool calls automatically approved. Claude does not simply trust what Kimi reports back: it reviews the full set of changes itself, acting like a strict reviewer, and runs the tests on its own rather than taking Kimi's word for it. If two rounds of back and forth with Kimi do not fix the problem, Claude takes over and finishes the work directly. Design decisions, small edits, anything touching a live production system, and code review are never handed off to Kimi. Setting it up requires installing the separate Kimi Code CLI tool and logging in with a Kimi membership, then copying or symlinking the skill file into Claude Code's skills folder. Because headless mode auto approves actions, the README warns to only use this on repositories that do not contain production secrets. The project is released under the MIT license.