Building a $0-marketing app with Claude Fable 5

UPDATE (July 15): The experiment left the lab as Kokoro is now live on the Japanese App Store: https://apps.apple.com/jp/app/id6788245039. The $0-marketing clock officially starts today; follow-up post with real numbers (good or bad) coming in ~6 weeks.

---

I ran an experiment: can an AI agent take a product from market research to a submitted App Store app, solo, with a hard $0 marketing budget? Not "cheap marketing." Zero. The app has to be discoverable or die.

This is the build log. The app — こころ (Kokoro), an anonymous AI self-care companion for Japan — is sitting in App Review as I write this, so no revenue numbers yet. What I can share is the method, because the method is the interesting part.

The setup

Everything ran through Claude Code with Fable 5 (Anthropic's new Mythos-class model). Not as a code autocomplete — as the other half of the team. It ran the market research pipeline, made the competitive analysis, wrote the launch plan with kill-gates, built the app, drew the mascot, generated the App Store screenshots, deployed the backend, and drafted the store listing. My job was decisions, accounts, and taste.

The starting point was a research doc with a small tool inside it: a JavaScript scanner that hits Apple's free, public APIs — top-grossing RSS charts (a revenue proxy), the iTunes lookup API (ratings), and the public review feeds. No Sensor Tower subscription. No API keys. Free data that most people don't realize exists.

Research before ideas

The agent re-ran that pipeline fresh across 6 storefronts: 1,798 top-grossing apps, filtered to wellbeing, then mined ~1,000 recent 1–2★ reviews from the interesting targets.

Three findings made every decision downstream:

1. Japan is the emptiest rich market. 24 wellbeing apps in 300 top-grossing slots (US: 118). This matters because a $0 marketing budget has exactly one growth channel: organic App Store search. ASO only works where the keyword space isn't saturated. The market choice wasn't a preference — it was forced by the constraint.

2. Paying users write your spec in their 1-star reviews. Japan's top-grossing CBT app: 38% of recent reviews negative. Verbatim: "the built-in AI feels a generation old — better to talk to free GPT-5.2." Another app's users said ChatGPT was more empathetic than the human counselors they were paying for. A third app's AI reframed a user's abuse story as "maybe the other side had circumstances too." Every one of those reviews became a product rule with a comment in the codebase pointing back to it.

3. Billing rage beats feature requests everywhere. The dominant 1-star theme in every single app we mined — Japan, US, Korea — was billing. Silent ¥19,000 annual renewals. Refusal to refund. Paywalls shoved at users right after emotional intake. Which means the cheapest differentiation available wasn't a feature at all: it was honesty. Pre-renewal notifications (a local push, no server needed), one-tap cancel, promises printed on the paywall itself.

The PM layer: kill-gates, not vibes

The plan isn't "build it and hope." Each stage has a pre-committed kill criterion:

• Gate 0 (before any code): does App Store autocomplete prove Japanese users search for this? (It did — including "認知行動療法 無料" and competitors' brand names.) Cost of being wrong: 2 hours.
• Gate 1 (before launch): do TestFlight testers keep journaling after a week, or drift back to ChatGPT? If they drift, fix the loop before launch — don't launch into a wall.
• Gate 2 (launch + 6 weeks): trial→paid ≥ 5%, rating ≥ 4.5. Under 2% after paywall iteration = the market said no, and the pivot (same codebase, different market) is already written down.

Pre-deciding the kill conditions is the discipline part. It's much harder to lie to yourself when past-you already defined failure.

Same logic drove the biggest scope decision: cloud backup with end-to-end encryption is the most expensive remaining feature, so it ships after the gates prove people retain and pay — not before. v1 is honest about it: local-only storage, stated plainly in the privacy policy.

Things the agent did that surprised me

• The mascot. It picked a capybara with a yuzu on its head — onsen capybaras are Japan's cultural shorthand for calm. Then it drew the SVG, wrote itself a design guide (geometry recipe, palette tokens, expression rules) so future art stays on-model, and rasterized the app icon with headless Edge because I didn't have proper design tooling installed.
• The App Store screenshots are HTML files rendered at 1320×2868 — same pipeline. The letter shown in screenshot #4 is a real output from the production backend, not lorem ipsum.
• Live debugging across three ecosystems in one evening: a space in a folder name breaking Xcode builds, Vercel dropping Node 18 while our blog engine emitted syntax removed in Node 22+ (fixed with a pre-build patch that rewrites assert { type: 'json' } to with { type: 'json' } inside node_modules), and Apple's screenshot-slot dimensions. None of these are in a tutorial.
• Provider-agnostic AI backend. The app never holds an API key: a Supabase edge function does, and AI_PROVIDER / AI_MODEL env secrets switch every user between Claude and OpenAI models with zero app releases. We switched providers mid-testing with a single dashboard edit.

The stack, for the curious

Expo SDK 57 (React Native), local-first AsyncStorage journal, Supabase Edge Functions as the AI proxy, RevenueCat for IAP (pay-first, no accounts — even payment is anonymous), PostHog for the gate metrics, and the whole build/sign/upload pipeline driven from the terminal via xcodebuild.

Total cash spent: the $99 Apple developer fee. Time: three days from reading the research file to "Waiting for Review."

What I'd tell you to steal

1. Apple's free APIs are a research goldmine. Top-grossing charts + review feeds = demand proof and a written spec, for free.
2. Pick the market your constraint forces, not the one you like. $0 marketing chose Japan for us.
3. Mine 1-star reviews of profitable apps. Angry paying customers are the highest-signal PMs on earth.
4. Make honesty a feature when the category is dishonest. It's the only differentiation that costs nothing and compounds.
5. Write your kill criteria before you write code.

The experiment continues — Gate 1 (Japanese TestFlight testers) is next, and the data decides from here. If you want the follow-up when real numbers exist, watch this space.

Built with Claude Code (Fable 5). Decisions, mistakes, and the capybara approval were human.