Case Study — Keeping One Voice Across 70,000 Words (An AI Content Pipeline That Actually Ships)

Client: ValorHub (self-owned)

Live proof:

What was shipped: A published business novel — 70,049 words, 29 chapters, real customers. Written by AI, produced by one person. Publishing quality, not "generated content."

Timeline: ~5 months, evenings and weekends.

The pattern:

Multi-agent content pipeline The Architect designs the Canon once, upstream. Every chapter then enters the runtime pipeline at Ghost Writer and flows through Editor, Polisher, and Screener. Drafts scoring below 0.95 loop back from Editor to Ghost Writer for revision — the Editor is the only gate that kicks work back. Architect designs Canon once · never writes prose produces once Canon — Bible · Beat Sheet · Framework the source of truth read by every specialist on every chapter score < 0.95 → revise New chapter 1 of 29 input Ghost Writer drafts chapter reads chapters 1..N-1 Editor 6-dim score gate: ≥ 0.95 Polisher catches LLM tells phrase blocklist Screener final AI check rhythm · cadence Published on-brand ✓ Chapter that scored ≥ 0.95 — flows to publish Chapter below threshold — Editor sends it back to Ghost Writer with a dimensional score
The Architect designs the Canon once, upstream. Every chapter then enters the runtime loop at Ghost Writer. The Editor is the only stage that can kick work back — and it does so with a specific dimensional score, not a vague "try harder."

Why a solopreneur running an AI content op should read this

You bought an AI writing tool. For two weeks it wrote your posts in your voice, your facts, your tone. Delightful. Time saved.

Two months in, every third post opens with "In today's fast-moving landscape…" One blog post says you serve 12 clients; the next says 20. Your website mentions three services; your last newsletter mentioned four. Nobody on your team noticed until a returning customer asked, mildly confused, which of the two numbers was correct.

This is drift. It is not a bug in your AI tool. It's what every AI writing setup does when nobody scores the output — the AI slowly forgets what your brand actually says. Facts blur. Voice smoothes. Repetition creeps in.

I killed drift at the hardest possible scale: a full-length business novel, 29 chapters, on Amazon. If the pipeline works there, it works for your blog + LinkedIn + newsletter.

The pipeline

Two roles: one planner upstream (before any writing), four specialists in a loop (every chapter).

Architect (upstream, once). Never writes prose. Produces the Bible — every character, every fact, every rule the book makes to the reader. Saved as a file. In business terms: your brand book + product truth-file + style guide, merged into one document an AI can read.

Then, for every chapter, four specialists run in order:

  1. Ghost Writer — drafts the chapter. Before writing, reads the Bible plus every chapter that came before.
  2. Editor — scores the draft on six dimensions (voice, plot, teaching, character, continuity, freshness). Anything below 0.95 goes back to the Ghost Writer with a specific dimensional score. The Editor is the only gate that can kick work back.
  3. Polisher — deletes AI tells ("a shiver ran down my spine," "at the end of the day," "navigating the landscape of"). Every rule was written after seeing the mistake in an earlier draft.
  4. Screener — last check. Sentence rhythm, cadence. Catches anything that still reads machine-written.

One planner, four specialists, one gate. Human review only when the Editor fails the same chapter twice — that's the signal to update the Bible, not to lecture the writer.

Why not one big AI doing all of it

Every client asks first. I tried it first too. Around chapter 4, the one-AI setup stops writing to the Bible and starts writing to its own last chapter. It pattern-matches against yesterday's output instead of the source of truth. Voice drifts. Character names get mangled. By chapter 10 the protagonist speaks with the wrong accent.

Give each stage a fresh AI with a narrow job and a clean read of the Bible, and drift stops. Five agents beat one because each starts from zero every time.

The proof

Read Chapter 1 free at valorhub.eu. Buy the rest on Amazon. Pick chapter 2, chapter 15, and chapter 28 at random. Check that the character sounds the same, the fictional company's facts match, and the recurring motifs accumulate weight instead of getting name-dropped and forgotten.

A pipeline that drifted at chapter 5 could not have shipped 29.

What working with me looks like

Bible-only kick-off — €1,200 (~2 weeks). For a small operator running one or two channels with an off-the-shelf AI (ChatGPT, Claude, Jasper — same shape for all). I build you the Bible; you paste it into whatever AI you already use, at the start of every session. One revision round included. Deliverable: the Bible file (yours forever) + a 15-minute call teaching you how to feed it. That's 80% of the drift fix for 10% of the pipeline price.

Build — €10,000–€18,000 (3–5 weeks, one channel). Full pipeline: your Bible, three specialist AI roles, a scored gate between them, a deploy runbook. Your team runs it after handoff. A marketing team paying a freelance writer €1.5k/month for 8 posts that drift by week 4 spends €6k/quarter on content that half-works. One pipeline build replaces that with scored, consistent output your team runs for the cost of API tokens.

Contact: info@valorhub.eu · linkedin.com/in/maksymdonets · Munich (CET)