You have an idea. Maybe it came to you in the shower. Maybe it's been in a note on your phone for two years. Maybe it's one sentence you've never finished.

Type it in. That's it. The system takes over.

8 AI agents run in sequence — researching your genre, forging a premise with a structural irony engine, writing every chapter, scoring each one independently, revising anything below the quality gate, and packaging the result for publication.

You approve 3 times. Everything else is automatic.

No writing experience required. No prompt engineering. No creative blocks.

Quick start

Step 1 — Get Claude Code

Download it free at claude.ai/code. It runs in your terminal.

Step 2 — Install the agents

# macOS / Linux
git clone https://github.com/felipelobomotta-blip/best-seller-studio
cp best-seller-studio/agents/*.md ~/.claude/agents/

# Windows (PowerShell)
git clone https://github.com/felipelobomotta-blip/best-seller-studio
Copy-Item best-seller-studio\agents\*.md $env:USERPROFILE\.claude\agents\

Step 3 — Give it an idea

Open Claude Code and type:

I have an idea for a book: [your idea here]

That's it. The book-orchestrator agent takes over and runs the whole pipeline.

What happens when you run it

Here's the exact sequence:

┌─────────────────────────────────────────────────────┐
│  YOUR IDEA                                          │
│  "what if a journalist covering the 2026 World Cup  │
│   discovers an alien contact cover-up?"             │
└──────────────────────┬──────────────────────────────┘
                       │
                       ▼
┌─────────────────────────────────────────────────────┐
│  PHASE 1 — Market Research                          │
│  • Finds 5–10 comparable published books            │
│  • Reads what readers hate about them (1-star       │
│    reviews are gold — that's your gap)              │
│  • Maps what's missing in the market                │
└──────────────────────┬──────────────────────────────┘
                       │
                       ▼
┌─────────────────────────────────────────────────────┐
│  PHASE 1.5 — Premise Forge  ⬅ the secret weapon    │
│  • Your raw idea becomes 5 structural variants      │
│  • Each is scored on 6 dimensions                   │
│  • The one with the strongest irony engine wins     │
│  • Floor required: 8.0/10                           │
│  • Rule: your idea is ELEVATED, never replaced      │
└──────────────────────┬──────────────────────────────┘
                       │
                       ▼
┌─────────────────────────────────────────────────────┐
│  PHASE 2 — Foundation & Outline                     │
│  • Character profiles (with contradictions baked in)│
│  • Full chapter map                                 │
│  • Every story beat intentionally subverted         │
│  • Voice DNA established                            │
└──────────────────────┬──────────────────────────────┘
                       │
             ╔═════════╧══════════╗
             ║   CHECKPOINT 1     ║  ← YOU APPROVE
             ║  "here's what we   ║
             ║   built and why"   ║
             ╚═════════╤══════════╝
                       │
                       ▼
┌─────────────────────────────────────────────────────┐
│  PHASE 3 — Chapter Loop (repeats for every chapter) │
│                                                     │
│  write → polish dialogue → sharpen hooks →          │
│  inject chaos → score independently                 │
│                                                     │
│  ✓ Score ≥ 8.5? → next chapter                     │
│  ↺ Score 7.0–8.5? → polish loop (max 5×)           │
│  ✗ Score < 7.0? → escalate to full revision        │
└──────────────────────┬──────────────────────────────┘
                       │
                       ▼
┌─────────────────────────────────────────────────────┐
│  PHASES 4 & 5 — Full Review & Revision              │
│  • Evaluates entire manuscript as one unit          │
│  • Re-attacks any weak chapter                      │
│  • Target: CVI-Launch ≥ 9.0                        │
│    (CVI = Commercial Viability Index — our          │
│     "Gone Girl tier" bestseller proxy)              │
└──────────────────────┬──────────────────────────────┘
                       │
             ╔═════════╧══════════╗
             ║   CHECKPOINT 2     ║  ← YOU APPROVE
             ║  "every chapter    ║
             ║   score shown"     ║
             ╚═════════╤══════════╝
                       │
                       ▼
┌─────────────────────────────────────────────────────┐
│  PHASE 6 — Delivery Package                         │
│  • Logline & back-cover blurb                       │
│  • Query letter (ready to send to agents)           │
│  • Formatted ebook + print files                    │
└──────────────────────┬──────────────────────────────┘
                       │
             ╔═════════╧══════════╗
             ║   CHECKPOINT 3     ║  ← YOU APPROVE
             ║   your book        ║
             ╚════════════════════╝

The quality gates

This is what makes Best Seller Studio different from just asking ChatGPT to write a book.

Gate 1 — Premise must score 8.0+ before a word is written

Every bestseller has the same structural DNA: a contradiction that generates conflict automatically. Gone Girl works because the "perfect marriage" premise immediately contradicts itself. Breaking Bad works because "chemistry teacher becomes drug lord" is a contradiction that escalates by its own logic.

The Premise Forge scores your idea across 6 dimensions:

The floor IS the score. A premise that scores 9/9/9/9/9/4 is a 4, not an 8.2 average.

Your raw idea is always Variant 1 (scored honestly as a baseline). If it already hits 8.0, it wins — the system doesn't forge for the sake of forging. Variants 2–5 are alternatives using different irony engines.

Gate 2 — Every chapter scores 8.5+ before the manuscript advances

The evaluator that scores your chapters is a separate agent that never wrote any of them. No self-grading.

It runs:

• 7-dimension Genesis Score (floor is the score)
• 20-pattern anti-AI scan (clichés, hedged language, fake profundity, on-the-nose dialogue...)
• Simulation of 4 reader types reading independently
• "Would you remember this chapter tomorrow?" test

Chapters between the hard floor and the 8.5 target enter the Polish Loop: the evaluator writes a surgical work order naming the exact failing dimensions, and the editor fixes only those. The rest is untouched.

Anti-inflation protection: each cycle can only add +0.5 maximum. You can't shortcut the gate.

Gate 3 — All chapters 8.5+ AND CVI-Launch ≥ 9.0

CVI-Launch is our best-seller potential proxy. 9.0 is "breakout potential, Gone Girl tier" — the manuscript has a word-of-mouth mechanism built in structurally, not just good prose.

The manuscript only reaches packaging when both gates pass.

The agents

All agents are plain markdown files in agents/. Read them directly to see exactly what each agent does.

Real examples

These ideas were run through the pipeline:

"A memoir-style essay about growing up between two cultures, not belonging to either"
→ Protocolo Não Encontrado — strong external response on early readers

"What if the world's top chess players are actually running a secret intelligence network"
→ Age of Aquarius — high internal Genesis Score after iterative evaluation

"A journalist covering the 2026 World Cup discovers an alien contact cover-up"
→ THE LAST COVENANT — currently in production, first V4.2 benchmark run

Cost

Using Claude Sonnet 4.6 (recommended):

Claude Code runs on your Anthropic account via OAuth — no separate API key needed.

Requirements

• Claude Code — log in with your Anthropic account (OAuth). No API key needed.
• ~30 minutes unattended runtime per book (the pipeline runs autonomously)

Honest caveats

What the system guarantees: The manuscript scores 8.5+ on the internal quality gate before leaving the pipeline. No chapter ships silently below target. Every gate failure is surfaced to you explicitly.

What the system does not guarantee: A literal bestseller. Cover design, marketing, timing, and luck are outside the manuscript and outside this system. The quality gate attacks the word-of-mouth mechanism (retellability + CVI) because that's the lever we actually control.

About the 8.5 score: It's an internally-calibrated ruler based on comps that hit bestseller lists — not a certified external measurement. It's the best proxy we can build before a real audience reads the book.

Live benchmark

THE LAST COVENANT is the first book running the complete V4.2 pipeline (premise forge → chapter gate → exit gate).

Chapters 1–9 in Phase 5 revision. Chapters 10–14 being written through the new gate. When it's done, every chapter's Genesis Score, evaluator work orders, and CVI-Launch will be published in SHOWCASE.md — the first real-world V4.2 benchmark.

About

Built by Felipe Lobo — developer from Brazil.

6 months of iteration. 10+ books shipped. MIT licensed.

X / Twitter · LinkedIn · GitHub