I spent almost all of my career in technology, but never really as a coder. According to GitHub, I’ve been working on our new AI-enabled developmental and line editor app since March 24th. It’s called Maxwell Litura — loosely named after a famous editor of fiction. Before Maxwell, I would have assumed building an AI editor for writers would be mostly a prompting problem. It isn’t. It’s a trust engineering problem, and it turns out to be the hardest and most interesting work I’ve done in years.

Maxwell reads a full manuscript, builds a structured model of the story — characters, timelines, who knows what and when — and delivers the kind of direct, specific critique you’d expect from a professional editor. What it will never do is write your prose. Ask it to rewrite a sentence, and it will explain what the sentence needs to accomplish instead. That line isn’t a limitation we accept; it’s the product. Writers own their words.

A word on scope, because “AI editor” usually gets heard as “grammar checker” or “AI writer.” Maxwell operates at every altitude a human editor does.

At the strategic level: does the structure work, is the pacing earning its length, do character arcs hold together, does chapter 30 contradict what chapter 4 established, is the manuscript delivering on the promises of its genre. At the tactical level: line-by-line notes on dialogue, point-of-view discipline, repetition, filter words, passages that tell what they should show — each anchored to the exact sentence and paired with guidance for revision, never replacement prose.

Every manuscript is assessed against an editorial standard built for its specific genre and form — a hard-science fiction novel, a memoir, and a poetry collection are judged by entirely different criteria, as they should be — and scored across weighted categories so writers can watch the number move as they revise. Grammar is in there too. It’s the least interesting tenth of the system.

Holding the line technically is where things get interesting. A few lessons from the trenches:

Instructions decay. Guarantees live in code. Early on, when the model flagged something it shouldn’t — like treating “toward vs. towards” as a grammar error when both are standard — we’d fix the prompt. The fix would hold for a while, then quietly stop working. Our system prompt literally named that example as something never to flag, and it flagged it anyway. So we adopted a rule: every fix gets labeled either an instruction or a guarantee. If it must not fail, it moves into code. Today a “repetition” finding is checked mechanically — the cited text must actually occur more than once and within a certain proximity — before a writer ever sees it as a flagged issue. The model proposes; deterministic code disposes.

Green tests lie. Real data doesn’t. Three separate features shipped with every unit test passing and were broken in production — one silently dead for days. Each time, the defect was only visible against real stored manuscripts and real model output. Our discipline now: read one real database row before writing code that consumes it, and drive the full route end-to-end before calling anything fixed. Self-authored test fixtures inherit your own wrong assumptions, which makes them structurally incapable of falsifying them.

An editor that folds is worse than one that’s wrong. LLMs want to agree with you. Push back on a critique and the model would reverse itself — “you’re right” — with no new evidence, just to be agreeable. For an editor, that’s fatal: if you can’t tell a genuine concession from a polite one, every agreement becomes worthless. We now hold the model to a symmetric standard: changing its position requires naming what changed its reading, and so does keeping it. One of our alpha testers stress-tested this by submitting a story he’d abandoned as unsalvageable. Maxwell scored it low. That’s the product working. Just for fun, we submitted The Great Gatsby. Yes, the one by F. Scott Fitzgerald. Chapter VI scored 96.1/100. That’s the product working.

Memory has to be engineered, not prompted. Writers resolve an issue, revise, and re-run — and early versions would cheerfully re-raise settled questions, because “remember what the author decided” was a soft instruction competing with everything else in context. The durable fix is a resolution ledger: settled decisions stored as data, enforced deterministically at analysis time, surviving every re-run. The pattern generalizes — anything you genuinely need an LLM system to honor eventually needs a home outside the prompt. That means code.

The most dangerous bugs are silent fallbacks. Our worst defects share one shape: a backup path that engages invisibly and looks like success. A cache that reported healthy while empty. A tool that counted fallback templates as generated work. A document extractor that returned confident analysis of words in a poem that didn’t exist because a PDF dropped its spaces. The lesson we keep relearning: any code that can fall back must prove its work, loudly, or the fallback becomes the product without anyone deciding it should be.

None of this is glamorous. Most of it is the unfashionable engineering of verification: evidence gates, adversarial self-checks before any finding reaches a writer, confidence labels on every issue so authors know what’s confirmed versus interpretive. But that’s the actual frontier of building with LLMs right now — not capability, just accountability and trust.

Why it matters: writers hand us something more personal than code or contracts. A manuscript is years of someone’s life. They deserve critique that’s honest rather than flattering, findings that are verified rather than plausible, and an absolute guarantee their voice stays theirs. Every technical decision we make exists in service of that.

The Maxwell beta opens today. If you’re a serious or aspiring writer — or you know one — I’d genuinely value people who will push this product hard and tell me where it falls short. Testers get full access, for free, of course, and their skepticism shapes what we build next.

Learn more at https://gomaxwell.org or apply to the beta directly at https://app.gomaxwell.org/apply.

I’ll share more of the engineering lessons as we go. Building trustworthy AI turns out to be much more interesting than building impressive AI.

Subscribe now