Blake Anderson open-sources ai-app-playbook, using a Coding Agent to build consumer-grade apps on a fixed stack…

Building a Profitable Consumer AI App with a Coding Agent on a Fixed Stack

@blakeandersonw's open-source "ai-app-playbook" is an operating procedure for coding agents to execute: human and machine split the work to go from product selection → build → launch → first-month profitability; the tech stack is fixed to the full Cloudflare suite, with an explicit long list of "don't use." https://t.co/G1zD5Yikef

What it is - two-layer structure: · Human-facing chapters playbook/00–13: from finding a niche and using agents to write code, to backend, database, login, models, client, payments, data, design, retention, distribution, cost, and compliance. · Agent-facing AGENTS.md: Phase 0–14, each step marked [AGENT] / [HUMAN]; before starting there's Gate A (only start if the score is high enough, there's demand signal, and unit economics make sense). The repo also includes templates/ (Worker, Better Auth, wrangler, app-level AGENTS.app.md). Usage is straightforward: clone, and in an empty directory point Claude Code / OpenCode at AGENTS.md.

Product direction By default, build consumer-facing products that are "useful to people and people are willing to pay for": fitness, focus, learning, calm, etc. A one-sentence template: For [who], [App] turns [input] into [result], completed in seconds. Seven-dimension scoring (pain, frequency, demonstrability, AI is core, willingness to pay, reachability, low risk); only consider building if ≥25/35 with no single item scoring 1. Explicitly avoid: generic chatbots, capabilities platforms will soon make free, diagnostic advice as the core, two-sided markets without supply, and grinding enterprise sales through the app store.

Revenue assumption (the author states this is a benchmark, not a target) Default about $30/year + trial. Rough funnel: downloads × 30% start trial × 30% convert to paid × $30 ≈ $2.70 per download (about 9% of downloads become paying users). The cost table is marked with 2026-10-26 pricing; the author claims their own stack (Cloudflare + Superwall + Tinybird + DeepSeek, etc.) costs about 2.5% of revenue at all scales (excluding store fees and ad spend). In the comparison table, Vercel, Supabase, Clerk, frontier large models, etc. are marked as "Popular," with large multiple differences.

Tech stack (mandatory) and anti-stack · Use: Workers (plain fetch, no framework) + D1 pure SQL (at scale, PlanetScale via Hyperdrive) + R2/KV/DO/Queues, etc.; Better Auth; AI defaults to OpenRouter → DeepSeek (via AI Gateway), at high volume can use https://t.co/7pS44wKG7o / RunPod; payments Superwall (client) + Stripe (Web); analytics and errors Tinybird; client: SwiftUI, RN or Kotlin, React + Vite; package manager bun. · Don't use: Vercel, Netlify, Lovable, Bolt, v0, Firebase/Supabase, Clerk/Auth0, Sentry/PostHog, Next.js, Hono, Drizzle/Prisma, and a long list of others. The principle is: if the stack can't do it, stop and ask a human; don't secretly add dependencies.

Phase rhythm (summary) After passing Gate A: Brief → accounts and repo → backend and database → login (anonymous first) → AI (streaming, timeout fallback, cost accounting, evals) → Web / iOS / Android → payments and entitlement verification → funnel events into Tinybird → onboarding polish → retention (push, streaks, email) → launch compliance → drive the first ~1000 downloads (short video, creator platforms SideShift/Trybe, DMs) → one funnel experiment + one product fix per week. Humans handle accounts, on-device signing, releases, and production; agents don't touch production data or production deployments.