🌶️ kimchi

Turn a raw idea into build-ready docs your AI can just execute.

Part 1 of 2 — kimchi decides what to build. 🎅 SantaClaude builds it. It picks up the EPIC docs and execute.md that kimchi emits and runs them as small, independently verifiable micro-tasks — dispatching each to a fresh subagent and auditing the work before it counts as done. Plan it here, ship it there.

Having ideas was never the problem. Getting your AI to build one correctly is. You open Claude or Codex, type a vague prompt, and burn hours re-explaining yourself while it fills gaps you never specified and builds the wrong thing — because the real spec only ever lived in your head. A few frustrating hours later the idea joins the graveyard.

kimchi gives you the spec. You hand it a raw idea; it hands back a clean, easy-to-read set of build-ready docs:

• One doc per EPIC — user stories with business severity/complexity and a strict priority order.
• Locked decisions — tech stack, architecture, exact API contracts, security and deployment calls, all fixed so a build session never re-litigates them.
• execute.md — a self-driving handoff that lets Claude Code or Codex build the whole thing story-by-story, autonomously, across multiple sessions, resuming from trackers.
• A net tracker + pre-build audit — whole-product status at a glance, and a gate that catches flawed decisions while they're still a doc edit, not a rewrite.

No more prompt-wrestling. You give the docs to your AI and it executes.

How it gets there: instead of one agreeable assistant, kimchi puts your idea through 11 world-class expert personas (product, architecture, design, security, finance, a deliberately lazy staff engineer, and more) that grill you, counter you, and refuse to flatter you until every decision is nailed down. The grilling is the engine — the docs are the gold.

Raw ingredients in. Pressure and time. A sharp, buildable spec comes out. That's the name.

Direct prompting creates AI slop. kimchi hands your AI a spec it can actually build.

---

What it does

• Grills the idea. Summon experts one at a time; each interrogates your idea in their domain until ~9 of 10 hard questions have clean answers. No "great idea!" — real friction.
• No appeasement. Every persona recommends with reasons and tells you when you're hand-waving past a real problem.
• Compiles build docs. One doc per EPIC — user stories, business severity/complexity, priority order, locked tech decisions, API contracts, and a tracker.
• Self-driving build. An execute.md handoff so Claude Code builds story-by-story in priority order across separate sessions, resuming from trackers.
• Audits before you build. A pre-build audit gate catches flawed or contradictory decisions while they're still a doc edit — not a rewrite.

The personas

PH product-head · CF co-founder · ARCH architect · DG design-girl · TB tech-bro (lazy staff eng) · SM security-master · DO devops · FB finance-bro (market research

• the money) · LB legal-bro · COO operations & execution · AUD auditor.

Install

Works in Claude Code and Codex CLI — both read the same SKILL.md format.

Claude Code — plugin (recommended: auto-updates)

/plugin marketplace add chitransh-cj/kimchi
/plugin install kimchi@chitransh-cj

Update later with /plugin marketplace update chitransh-cj. kimchi ships as a single skill, so it invokes with the clean name /kimchi (e.g. /kimchi PH) — no namespace prefix.

Claude Code / Codex — clone + install script

git clone https://github.com/chitransh-cj/kimchi.git
cd kimchi
./install.sh          # auto-detects installed agents
# or force one:  ./install.sh claude | codex | all

• Claude Code → ~/.claude/skills/kimchi/ → invoke /kimchi
• Codex CLI → ~/.codex/skills/kimchi/ → invoke via /skills menu or $kimchi

Restart the agent after installing so it picks up the skill.

Codex — manual

Copy the skill into your Codex skills directory (personal or project-local):

git clone https://github.com/chitransh-cj/kimchi.git
mkdir -p ~/.codex/skills/kimchi
cp -R kimchi/plugin/SKILL.md kimchi/plugin/references ~/.codex/skills/kimchi/

Usage

Invoke /kimchi (Claude Code) or $kimchi (Codex). Arguments are the same either way.

/kimchi              # no argument — full product-development lifecycle, guided
/kimchi CF           # summon one persona by name or acronym (co-founder here)
/kimchi PH           # product-head — builds the EPIC → STORY → TASK breakdown
/kimchi generate     # compile the build docs (README, EPIC-*.md, execute.md)
/kimchi audit        # run the auditor over the docs (do this before you build)

Start with /kimchi and let it walk you through. Answer honestly — it's designed to catch you when you don't.

License

MIT