A Claude Code skill that routes implementation work-orders to Kimi Code CLI (Kimi k3) lanes — Kimi types, Claude thinks and verifies.
Claude keeps the judgment work: specs, architecture, orchestration, review, and the merge button. Kimi gets the hands-on volume: frozen-spec builds, mechanical migrations, test fills, bulk sweeps, CI fixes. Kimi's flat-rate membership makes it a cheap, high-throughput implementation lane while Claude spends its tokens on the parts that need judgment.
This is a port of @steipete's excellent codex-first skill (from agent-scripts) — same delegation doctrine, swapped from the Codex CLI to the Kimi Code CLI.
When a task reads as a work order (the spec is frozen, no design decisions left), Claude:
- Writes a full-context prompt to a temp file — goal, repo paths, constraints, non-goals, expected proof, output shape. Kimi starts with zero session context, so spec quality decides success.
- Runs
kimi -pheadless in the target repo (tool calls auto-approved, thinking to stderr, answer to stdout). - Reviews the full diff like a hostile PR reviewer, runs the tests itself — Kimi's claims are advisory.
- Iterates via
kimi --continue; after 2 failed rounds, takes over and does it directly. - Ships through normal repo protocol — Claude's hands on the merge button, never Kimi's.
Design work, tiny edits, prod operations, and review are never delegated.
1. Kimi Code CLI (skip if you have it — check with kimi --version):
curl -fsSL https://code.kimi.com/kimi-code/install.sh | bash
kimi login # device-code flow, needs a Kimi membershipDocs: moonshotai.github.io/kimi-code
2. The skill, into Claude Code:
mkdir -p ~/.claude/skills/kimi-first
curl -fsSL https://raw.githubusercontent.com/boringmarketer/kimi-first/main/skills/kimi-first/SKILL.md \
-o ~/.claude/skills/kimi-first/SKILL.mdOr clone and symlink:
git clone https://github.com/boringmarketer/kimi-first.git
ln -s "$(pwd)/kimi-first/skills/kimi-first" ~/.claude/skills/kimi-firstClaude Code picks it up automatically; invoke it explicitly with /kimi-first or just hand Claude a work-order-shaped task.
The CLI mechanics in the skill were live-tested against Kimi Code CLI v0.26.0:
kimi -pruns headless with tool calls auto-approved — answer on stdout (•-prefixed), thinking on stderr--yolorefuses to combine with-p(and isn't needed)kimi --continue -presumes the latest session for that directory with full context--continueis cwd-filtered — never run two concurrent sessions in the same directory- Headless mode auto-approves, so only delegate from repos without prod secrets in reach
- @steipete for the original codex-first doctrine this ports
- Built by @boringmarketer · boringmarketing.com — Search & AI visibility, in Slack