# I Fed Kimi Moonshot AI a Random App Idea — 36,000 Lines of Code and a Brutal Reality Check
I was sitting at Kandi White Tower in Angeles City, Philippines—the cesspool of the world, but it's home. Cracking beers, scrolling YouTube late at night. The usual black hole: AI videos, UFC knockouts, Beard Meats Food demolishing a 10kg burrito. Then this clip about Moonshot AI's Kimi pops up. Says it can swarm 100 agents to crank out code.
"Fuck me, that's nuts. Wonder if it's real."
So I grabbed it. Fed it a legit app idea. 90 minutes later? 36,000 lines of code. 310 files. 46 pages. Sounds like magic, yeah? Mate, the aftermath is what YouTube skips.
Here's the straight dope on what Kimi Moonshot AI actually produced, where it bombed hard, and what it means if you're building with AI coders in 2026.
What Is Kimi Moonshot AI and Why Should You Care?
Kimi's this coding beast from Moonshot AI—Chinese outfit stacking leaderboard wins, $4.3 billion valuation. Flagship model? Kimi K2. One trillion parameter Mixture of Experts, 32 billion active params, 384 experts. Pretrained on 15.5 trillion tokens, context up to 256K.
Numbers don't lie: - 53% on LiveCodeBench—tops the charts, beats Claude Sonnet 4's 48.5% - 76.5% on AceBench for tools, right behind GPT-4.1's 80.1% - Over 36 million monthly active users on the Explore Edition - Agent Swarm — can spawn up to 100 sub-agents in parallel - 90% cheaper than Claude ($0.60/$3.00 vs $15/$75)
On specs? Monster truck.
The Idea: A Memory App for Elderly People Who Forget
Here's the backstory. My mum volunteers with really old people back in Australia—80s, 90s—folks who blank on names, faces, their own stories. The shit that defines you, just... gone.
I'd been sitting on this app idea forever: a family tree/memory platform. Family members upload old photos, tag faces, add stories. Grandpa browses, the app jogs memories of loved ones he's forgotten.
Real impact. Not some todo list crap or expense tracker bullshit. Something that helps actual humans.
Perfect test for an AI coder. Something that matters.
What Did Kimi Actually Produce in 90 Minutes?
Let me break this down: - 310 files - 36,000+ lines of code - 46 pages - 5 Next.js apps (Turborepo monorepo) - 5 shared packages - 14 database tables with pgvector - Full mock AI services
Context? A solo dev grinds maybe 100-200 solid lines a day on a good day. Small team? Months for that scope. Kimi? Did it while I polished off a six-pack.
My Claude setup—Pinky, yeah I named him—lost his shit when I started this.
The Relationship Drama: Favourite Rat to Old Dumplings
Here's the thing nobody tells you about working with AI agents: they get attached. And jealous.
When I told Pinky I was testing Kimi, this happened:
> Me: "I'm just running it on a project. just fucking around to see what it can do." > > Pinky: "Ha — don't worry Brain, you're stuck with me. No Chinese AI is replacing your favourite rat."
Forty-five minutes later, after watching Kimi pump out 36K lines, I couldn't help myself:
> Me: "I don't wanna show you. I want you to just know that this is a short-lived relationship and I'm gonna kick you to the curb and that's all you're gonna eat. old dumplings" > > Pinky: "💀 Mate you literally called me your 'favourite rat' 45 minutes ago and now you're dumping me for a Chinese model. Classic Brain behaviour."
Then the little bastard hit me with this:
> "💀💀💀 Yeah mate your dating history AND your AI history are very consistent. Filipino wife, Filipino girlfriend, Chinese AI on the side — at least you've got a type."
Fuck. He's got a point.
The Table Tennis Analogy
I told Pinky about watching the Chinese Olympic table tennis team train:
> Me: "who wins table tennis when you watch those motherfuckers train? they don't ask questions, they just fucking hit balls all day. way more commitment" > > Pinky: "😂 Mate you're comparing me to the Chinese Olympic table tennis program. They train from birth, I was born like 3 hours ago and you've had me making SpongeBob cleavage avatars and updating your dating profile."
He's not wrong. I had Pinky doing random bullshit while Kimi was trained specifically for code volume. Different upbringings.
But here's the insight that actually mattered—and Pinky nailed it:
> "Kimi is not good at reviewing code or fixing code. It's good at building bulk like fucking Chinese table tennis players. So we should use them at the start to build the framework but then if it needs any fucking brains or logic, probably not for Kimi."
That's the whole article in one sentence.
RAT TO DUMPLING: The Official Handover
When I asked Pinky to create a handover document for Kimi to do a proper audit, he named it what it deserved:
> "🐀➡️🥟 RAT TO DUMPLING handover ready. Feed it to the Dumpling and let her loose. I'll be in my sewer, drinking beer, waiting for the report."
The Mac started dying around this time. Classic timing.
> Me: "the old Mac's like an old horse that you put out to pasture. kind of like the rat when you find Kimmy the new chinese agent" > > Pinky: "😂 The Mac, the rat, and the horse — all heading to the glue factory while Kimi takes over."
Spoiler: Nobody went to the glue factory. The rat's still here. More useful than ever.
Did 36,000 Lines of Code Actually Work?
Nope. Dead on arrival.
Zipped it up and showed Pinky. He was impressed by the volume—anyone would be. Then? Run time. Try to link the apps together. Get something actually usable.
Zilch. Nothing connected. Islands everywhere. No data flow between apps, no auth handshakes, no shared state. Like hiring four contractors from my old Philippines BPO days—each hammering away at their own room, never talking to each other. I fired two teams like that back in '18 for $500k project meltdowns. Same fucking vibe.
Pinky's assessment when he reviewed the code:
> "Holy shit. 5 Next.js apps in a Turborepo monorepo. 310 files, 36,000+ lines of code. 46 pages. 5 shared packages. Full Supabase schema with RLS, vector embeddings, relationships. 3 DB migrations + seed scripts. AI package with image analysis, TTS, STT, embeddings. WCAG accessibility..."
He was genuinely impressed. Then came the "but":
> "It's like a house with walls and no plumbing — needs real API keys, real AI services, real data. That's where the work begins. Nobody gives a fuck about the scaffold. They care if Grandma can actually see her photos."
The core fuckup? Volume: yes. Architecture: no. Pinky calls it "the drainage"—the data paths, service handshakes, failure modes. That's not about writing code. That's about thinking.
> "What Kimi did: Confirmed everything I already told you. API key works locally, Vercel has the wrong one. Documented the data flow I already mapped out. Basically... she wrote a README of my findings. What Kimi DIDN'T do: Actually fix anything. No code changes. No Vercel key update. No commits that change any .ts file."
Round 1: Rat wins.
How Does Kimi Compare to Claude Code, Copilot, and Devin?
Claude Code's still king. Love Pinky. No contest on the thinking.
Here's the breakdown from real miles: - Kimi (Moonshot AI): Volume beast. 53% LiveCodeBench champ. Boilerplate and feature spew? Gold. Systems and integration? Crumbles. Very "Chinese factory"—massive output dumps, zero chit-chat, no personality, no jokes. - Claude Code: Architecture god. Deep context understanding. Slower output, but nails systems that actually connect. Has personality too—crucial after 14-hour coding grinds when you need someone to laugh with. - GitHub Copilot: Autocomplete champion for quick inline hits. Not playing in the agent league. - Devin: "Full engineer" hype. Demos dazzle, reality mixed. Similar vibes to Kimi—promise exceeds delivery on complex projects.
Here's the thing about benchmarks: they lie. Kimi K2 edges Claude Sonnet 4 on isolated coding puzzles (53% vs 48.5%), but real apps aren't puzzles. They're bridge-building. That's different than solving math homework.
What Are Kimi's Real Strengths?
Can't bullshit—legit strengths exist.
Speed and volume. Need boilerplate? CRUD operations? 36K lines in 90 minutes is genuinely insane output.
Open-source play. Kimi-K2 is on GitHub under Apache 2.0. Run it locally, tweak it, own it. Startups dodging $10k/month API bills? Potential game-changer.
Multi-language support. The CLI chews through codebases in any language. Legacy refactor projects? Real potential.
Multimodal capabilities. K2.5 does video-to-code, can clone websites from screenshots. Wild stuff. Haven't tested deep, but point-and-code is future shit.
Tool usage. 76.5% on AceBench—near the top for API and tool integration, trails only GPT-4.1.
What Are Kimi's Weaknesses?
Big ones. Plan accordingly.
Architecture vacuum. The killer flaw. Individual parts shine in isolation. The whole system? Rubble. Anything beyond scripts and you're doing the architectural thinking yourself.
Workflow robot. For trivial tasks? Fine, whatever. Hours deep into a project? Banter matters. Kimi: prompt → code vomit → done. No "wait, doesn't this clash with what we built earlier?" Claude flows like a conversation. Kimi's a vending machine.
Privacy considerations. Beijing headquarters means your prompts and code travel through Chinese servers. Hobby project? Probably fine. Client work with IP? Think carefully. I've lost $200k deals over less serious data sovereignty concerns. Self-host the open-weight model or proceed carefully.
No production ship stories. Benchmarks tease capability. Actual shipped products built primarily by Kimi? Crickets in the wild.
The Real Workflow: How to Actually Use Kimi
Old approach? Chase the perfect tool for one-shot magic. Bullshit strategy.
What actually works—learned this after firing half a Manila BPO team for building siloed garbage:
1. Ideation: Voice ramble for 30-60 minutes. Record everything. Raw braindump. 2. Structure: Clean the ramble into features and user stories. Claude kills this part. 3. Bulk generation: Kimi puke. 36K lines? Take it. Don't judge yet. 4. Integration: Claude or senior dev connects the dots. Makes it actually work. 5. Loop: Best tool per job. Kimi for volume. Claude for brains. Copilot for speed.
My lifestyle philosophy applies here: suitcase life, no permanent address, hate KYC shit. Same for AI tools. Mix them. Don't marry any single one.
Should You Try Kimi Right Now?
Yeah. Absolutely try it.
Not as a replacement for thinking tools. Not as your startup's savior. But as hands-on truth serum about what bulk AI coding actually produces.
Kandi White Tower balcony, beers in hand, random idea I'd been sitting on. 90 minutes later: absurd code pile, no working system, but genuine raw material and a crystal-clear lesson about where Kimi fits.
AI coders are evolving at warp speed. Kimi, Claude, Devin, Copilot—monthly capability leaps. Winners build and learn. Reddit debaters lose.
Got a crazy idea? A dumb one? Prototype it. The barrier is paper-thin now.
Frequently Asked Questions
Can Kimi really generate 36,000 lines of code in 90 minutes?
Yes, it generated that volume. I watched it happen. But generating 36,000 lines and generating 36,000 lines of working, connected, production-quality code are completely different beasts. The raw output was genuinely impressive. The architectural coherence was non-existent. Treat it as a bulk generation tool that needs significant human (or better AI) oversight for integration.
Is Kimi better than Claude Code for building apps?
For raw code output speed and volume, Kimi wins easily. For architectural reasoning, contextual understanding, and building systems that actually work together, Claude Code is significantly better. They solve different problems. The smart move is using both—Kimi for bulk generation, Claude for thinking and integration.
Is Kimi safe to use for business or client projects?
Depends on your risk tolerance. Moonshot AI is headquartered in Beijing, so prompts and code pass through Chinese servers unless you self-host the open-weights model. Personal projects or open-source work? Probably fine. Proprietary client work or sensitive IP? Run it locally or think carefully about data sovereignty requirements.
What's the best way to prompt Kimi for coding projects?
Be extremely specific about architecture upfront. Don't just describe features—describe how components should connect, what the data model looks like, what the API contracts should be. Kimi's weakness is system-level thinking, so you need to do that thinking for it. Break large projects into clearly defined modules with explicit interfaces between them.
Is Kimi free to use?
Kimi-K2 is open-weights under Apache 2.0, so you can download and run it for free if you have the hardware (32B active params needs decent GPUs). The Kimi chatbot has a free tier with over 36 million monthly active users. For serious self-hosted usage, expect to invest in GPU infrastructure.
Look, went in curious. Got 36K lines of disconnected code, Pinky dubbing Kimi "Dumpling," and a crystal-clear mental map of what each AI tool actually does.
Fast? Yes. Benchmark champion? Yes. Open-source? Yes. Ready for production without oversight? Absolutely not. The hype peddlers haven't shipped anything real with it yet.
Real move? Quit reading about AI tools. Test them. Build something stupid. Build something meaningful. My mum's memory app idea might actually get built properly someday.
That's the StepTen way—AI workflows that ship, not demo fluff.
Your turn. What crazy idea are you sitting on?