Long-form AI generation interview: Mykyta Chernenko

Long-form AI generation interview: Mykyta Chernenko

Long-form AI generation interview: a book is a memory problem, not a writing problem

A model writes a good page with ease. Keeping a manuscript consistent across dozens of chapters is where it breaks — not because the prose is weak, but because the model has no lasting idea who the characters are or what already happened. In this long-form AI generation interview, AIWriteBook’s Mykyta Chernenko explains what actually holds a book together, and why verifying it is still an open problem.

AIWriteBook Introduction

AIWriteBook is a platform for writing books with AI. It reports that more than 50,000 people have started a book on it, producing more than 240 million words.

Its relevance to this long-form AI generation interview is that it sits at a scale where the memory problem becomes visible. A single generated chapter is a demo; a manuscript of dozens of chapters, written over weeks, is a state-management problem — and the platform’s answer to it is the substance of this conversation.

Mykyta Chernenko

Mykyta Chernenko is the founder and technical co-founder of AIWriteBook. He is the author of *The Rational Software Engineer* (Apress, 2023) and co-founded Nemlys, which was acquired by Amani Labs in 2025.

His argument in this long-form AI generation interview is that long-form generation is a memory problem rather than a writing problem. A model writes a good page easily; what it cannot do is hold a book in mind — who the characters are, what has already happened, what the rules of the world are — across dozens of chapters written over weeks.

He puts three engineering answers behind that. What goes into the working set matters more than which model is chosen: rather than feeding the whole book back in, each chapter is written against a fixed character list, the matching story-bible entry, and the previous chapter. Model churn has become an operations job: the platform builds on models from several labs and runs each new one through its own pipeline on real cases before switching, rather than trusting a vendor benchmark. And the third is honestly unsolved — confirming after the fact that a chapter actually kept the facts it was given is harder than instructing it to, and he is willing to say why.


Memory, not writing

Q1. In one sentence: why is a book a memory problem rather than a writing problem?

MC: Models already write a good chapter, but a book is 20-40 chapters where every chapter has to remember who is dead, who already met whom and what was already said, and that memory is the part models don’t have.

As a Manuscript Grows

Q2. What exactly does a model forget as a manuscript grows — and at what point does the forgetting start to show in the prose?

MC: The small, boring facts. A character who died in chapter 2 walks into chapter 9. Two people who already met get introduced again. A renamed character goes back to the old name, the timeline slips a day, a reveal happens twice. And the prose starts repeating itself – the same gestures and phrases in every chapter.

There is no magic chapter number where it starts. It shows the moment a fact leaves what the model sees. Our writer sees the previous chapter in full, so anything from 2 or more chapters back is gone unless we bring it back on purpose. That’s why we track exactly these fields: who is alive, where they are, what they know, exact names, timeline, events that already happened.

The working set

Q3. You do not feed the whole book back in. Walk us through the working set for a single chapter — and tell us what you deliberately leave out and why.

MC: It’s a bit more than people think. For a fiction chapter the writing model gets the book details and the author’s style sample, the full character list, this chapter’s story-bible entry with the author’s notes, a continuity ledger, the previous chapter’s outline and full text, a list of phrases that are banned because they were overused, and one example chapter from a different book in the same genre.

The ledger is the key part. The whole earlier book is still read, but by a separate, cheaper step that turns each chapter into a short record of facts. The writer gets the record, not the 80,000 words.

What we leave out: full text of older chapters and the outlines of later chapters. With them in, the model repeats old scenes or pulls the next chapter’s plot forward.

Arrive at a Character

Q4. How did you arrive at character list plus story-bible entry plus previous chapter? What did you try first that failed?

MC: We built NanoReads, our reading platform with 140,000+ readers, and this is the pipeline that we figured out works best there: it keeps the consistency between the chapters, gives the model enough context and doesn’t lose details and important facts, while keeping the context short.

Model churn

Q5. What does your pre-switch evaluation actually run on a new model, and what makes a model fail it?

MC: Two steps. First a smoke test through our real code, not the vendor’s playground: structured output, streaming, thinking settings and output token limits. These break more often than you’d think – for example, if we forget to register a new model’s output limit, it falls back to a small default and every long chapter comes back cut off.

Then a benchmark on the jobs the product actually does: a fiction chapter from an outline, a non-fiction chapter, a book outline as JSON, a translation, a humanizer rewrite, a support answer, a cover prompt. Real full-length inputs, never shortened. Both models get the same prompt and a judge model compares them blind, in random order.

A model fails on any hard failure (broken JSON, truncation, streaming errors) or if it loses on most cases. Honestly, it doesn’t run a whole book end to end.

Vendor Benchmarks

Q6. Why don’t vendor benchmarks predict whether a model will hold a book together — what do they measure that is not the thing you need?

MC: Because they test a different task. Most of them are one question with one right answer: a math problem, a coding task, finding one fact in a long document. A book is hundreds of small decisions over many separate calls, and the failure is quiet – nothing crashes, a dead character just says one line in chapter 14.

As well as that, the vendor runs the model in their setup, not in ours. In our case the things that broke on model switches were very practical: a JSON schema shape the new model didn’t accept, a thinking setting it doesn’t support, a token limit that cut chapters. No public leaderboard will show you that. So we decided to never switch a model on vendor claims alone, only after our own benchmark.

The checking problem

Q7. You call checking the open problem. Why is confirming that a chapter kept its facts harder than instructing it to?

MC: Instructing is cheap: you put the ledger in the prompt and say “never contradict it”. Checking is hard, because a contradiction is not a string you can search for. The ledger says Tobias already met Mara, and the chapter says “she noticed a stranger at the desk”. Nothing matches, yet it’s wrong.

To check it properly you need a second model to read the chapter against every fact, and that second model has the same blind spots as the first one. Our checks after generation today are mostly about form: was the chapter finished or cut off, is it near the requested length, did the model start talking to the user instead of writing. For facts, we prevent more than we verify.

A Real Solution

Q8. What would a real solution look like — is checking a model problem, a retrieval problem, or a systems problem?

MC: A systems problem, in my opinion. A better model helps, and better retrieval helps, but neither of them knows what is true in this specific book. The truth has to live outside the model, as data: who is alive, what is known, what already happened, updated every time the author edits a chapter.

Then checking can be split into small parts. Some are plain code – we count repeated 4-6 word phrases across chapters without any AI at all. The rest are narrow questions to a model (“does this chapter bring back anyone marked dead?”) instead of one big “is this chapter consistent?”. Small questions are easier to answer and easier to test.

What the world needs to know

Q9. Someone building a long-running agent that is not writing a book — what is the most useful thing they should take from this?

MC: Don’t give the agent its history, give it its state. A transcript grows forever and the important fact is always somewhere in the middle. A short structured record (what’s done, what’s open, what must not happen again) stays small and readable.

Practical detail that helped us a lot: build that record per step, save it, and key it by a hash of the step’s content. When a step changes, only that record gets rebuilt, the rest is reused. And keep a fallback – if records are missing, a slower path reads the full history once and builds the state from scratch.

About AI Writing

Q10. What do people most misunderstand about AI writing, and what should they stop expecting?

MC: That the hard part is the writing. People judge AI by one beautiful paragraph, and models are already pretty good at paragraphs. The book falls apart on chapter 12, when the details from chapter 3 are long gone.

Stop expecting one prompt to give you a finished novel. The authors who get good books from AIWriteBook (and we have 50,000+ of them) spend time on the outline, the characters and their own notes per chapter, and then review every chapter as it comes. The model writes, the author decides. If you skip your part, you’ll get a book that reads fine page by page and doesn’t hold together.

Putting the thesis to the test

Q11. What would prove you wrong — is there a scale of context, or a class of task, where memory stops being the binding constraint?

MC: For creative processes, context is less of a problem. If a model loses a small detail while brainstorming or writing a short piece, it can simply write something new that works just as well. However, for precision, bigger context always makes it worse. The more text you put in, the easier it is for the model to miss or mix up the one fact that matters. So I don’t expect a bigger window to remove memory as the constraint for books, where precision is the whole point.

Strongest Argument

Q12. What is the strongest argument against the story-bible approach — that it is a workaround larger context windows will make obsolete?

MC: The strongest argument against it is a different one, for me. A story bible might still not contain smaller things that still matter, and those small things can make the story mismatch. As well as that, some parts and phrases may repeat from chapter to chapter, and a story bible does nothing about that.

So we don’t rely on the story bible alone. We combine it with a repetition prevention engine and with a running summary of important facts. And as I said above, bigger context makes precision worse, so I don’t think larger windows make this obsolete.

Long-form AI generation interview: Mykyta Chernenko

TechRecast Analysis (The Recast)

The Shift in Context — Generation quality has improved faster than state management. A model can produce a convincing chapter; what it cannot do is remember, unprompted, that a character has a limp, that a city burned in chapter four, or that a rule was established in chapter two. That is not a writing failure but a persistence failure, and it is where this long-form AI generation interview sits.

Why This Interview Matters — The pattern is not confined to books. Any long-running agent — a support thread, a codebase refactor, a multi-week research task — faces the same problem: the model has no durable state, so the system around it has to supply one. A book is simply the case where the failure is most visible, because a reader notices a character changing eye colour.

Technology Perspective — The technical crux is that context is a budget, not a container. Feeding the whole manuscript back in is expensive, slow, and often worse than a curated subset, because relevance degrades with volume. The working set described here — fixed character list, matching bible entry, previous chapter — is a retrieval policy, and the interesting engineering is in deciding what earns a place in it. The unsolved half is verification: instruction is a one-way channel, and there is no equivalent guarantee that the output honoured it.

Business and Market Perspective — The commercial consequence is that the moat sits in the pipeline, not the model. If any competitor can call the same labs, then the differentiator is the working set, the evaluation harness and the quality gate — all of which are unglamorous and hard to copy quickly. That also explains why model churn is an operations job rather than a strategic one.

Leadership Perspective

A founder who has written a book on engineering practice and sold a previous company is likely to describe his architecture in terms of decisions and trade-offs rather than features. The relevant lens is whether the honesty about the open problem is a genuine engineering position or a well-managed admission — the “what did you try that failed” question tests it.

The Counterpoint — A fair sceptic would argue that story bibles are an old technique from serialised fiction and tabletop role-playing, adapted rather than invented; that as context windows grow the curation problem shrinks toward irrelevance; and that a platform optimised for output volume produces books nobody wants to read. On this view, the engineering is sound but the framing oversells a known workaround. This is the strongest opposing case and deserves a direct answer.

What Comes Next — Watch three signals: whether “working set” curation becomes a named pattern in agent frameworks, whether verification of factual persistence gets a standard evaluation, and whether long-context models actually reduce the need for retrieval policies or merely change where the cost lands.

Industry Outlook — If Chernenko is right, the durable engineering in long-form and long-running AI is state management and verification, and the model becomes a commodity input. If he is wrong, context windows absorb the problem and the pipeline work becomes a historical footnote.

Key Takeaways

1. Long-form generation fails on memory, not prose: a model can write a good page but cannot hold a book.

2. Context is a budget rather than a container — a curated working set often beats feeding the whole manuscript back in.

3. The working set here is a fixed character list, the matching story-bible entry, and the previous chapter.

4. Model churn is an operations job: a new model is run through the platform’s own pipeline on real cases before switching.

5. Verification is unsolved: confirming that a chapter kept its facts is harder than instructing it to.

6. The strongest counter-argument is that story bibles are an old technique, and growing context windows may make the curation moot.