How we write.

Here is what "LLMs don't sound human" actually means in practice. These are the specific failure modes we work around — and the ones we expect to keep working on for some time.

• Generic emotional description. Left alone, the AI reaches for "she felt a pang of sadness" when what's wanted is a specific physical or behavioural detail. Our style rules push for the specific; drafts that settle for the generic get rewritten.
• Essay-voice dialogue. AI dialogue often sounds like the author talking rather than distinct characters. We watch for every character sounding the same and rewrite until the voices separate.
• Structural forgetfulness. Across a long book the model loses track of earlier decisions. The plot and chapter outlines are maintained by the human editor, not the model, and the AI is re-briefed on relevant context each chapter.
• The most-likely next sentence. Left alone, an LLM reaches for the statistically likely continuation — which is also the blandest. Style rules and rejected drafts are how we counter this. It's the single most common reason a draft gets rewritten.

As of April 2026: one pen name live (Nico Reyes, techno-thriller), one novel published — The Navarro Switch, 28 chapters, first person, around 85,000 words, a Bay Area countdown thriller. The manuscript ran through multiple continuity-audit passes and a copy-edit pass before publication.

What's working. The planning layer holds. Plots, character arcs, and structural beats are settled before a word of prose is generated, locked into a continuity ledger, and cross-referenced through every later pass. Across the seven plot threads in The Navarro Switch, every inciting incident, climax, and resolution landed in the chapter the outline planned for. Character arcs match the planned arc summaries. Around seventeen technical specs — kill-switch current, firmware versions, signal propagation timings, and so on — stayed consistent across all 28 chapters.

What we're still getting wrong. The audit catches things, but they keep recurring. On Nico's manuscript the first pass turned up: timeline references that didn't add up, transit lines named that didn't serve the route, technical terminology drifting between chapters, and early-chapter conversations re-litigated rather than evolved when they came back later. Five fixes were applied and the audit skill itself was updated to catch the pattern earlier next time. Em-dash frequency is the open issue we still haven't beaten — manuscript average around 10.7 per 1000 words, roughly twice what the style rules call for.

What the next book will try. The Last American Summer by Hazel Quinn — a coming-of-age historical novel in planning. 1989, a small Midwestern town whose auto-parts plant is closing after a tariff rollback, four teenagers on the cusp of adulthood pulled into the community fallout. A complete genre departure from the Nico Reyes line. Different pen name, different voice, different structural problems. The point of this one is to find out which parts of the method are universal and which were Nico-specific.

The method at this level of detail is public. The prompts, style rules, and voice profiles that make it work stay with us. The same loop runs at every layer: the AI proposes, a human directs, and drafts are accepted, redirected, or reworked.

Most of the directing happens in planning. The AI suggests a spread of pen names across genres; we review the voices, reshape the weaker ones, and drop the ones that don't work. Book concepts arrive three ways: some are purely human ideas, some are AI-proposed sequels and series entries, some start as AI seeds pushed into shape through conversation. From there the AI develops plots, characters, settings, worldbuilding, and research, with human input on every pass. This is where a novel gets its spine.

Once the plan holds, the AI writes the prose chapter by chapter against the style guide and the plan. Separate AI review passes follow: a copy edit for grammar and rhythm, a continuity audit across timeline, character facts, and geography. Chapters are refined through those passes, sometimes lightly, sometimes close to a rewrite. If something conflicts with earlier material, the conflicting part is re-drafted. The human spot-checks at key points: early chapters to catch the book going off-track, final chapters to catch the landing.

Readers come in after publication. They read the finished book and tell us how it landed: what worked, what broke, what rang false. Feedback shapes the next book. It can also put a current one back through the loop: strong feedback, or a hallucination we missed, triggers a new edition. The loop is the product.