Nothing special - I just asked an AI (ChatGPT) about it, and it confirmed it's an issue... the TLDR, it's probably related to how you form the prompts, and is self-reinforcing when it happens, like a feedback loop when a microphone gets too close to the speakers. But there are ways to hold it off.

Me:
What causes the staccato responses from Chat GPT when writing a story chapter once the context gets above around 12K tokens?

Chat:
Yes. What you are describing is a real long-context failure mode, but ~12K tokens is not a published ChatGPT limit. OpenAI's current GPT-5.6 models have much larger nominal context windows, so what you're seeing is almost certainly quality degradation from context composition, not the model running out of context. OpenAI Platform

For fiction, it often manifests exactly as you describe: prose becomes short, declarative, rhythmically repetitive, one-line paragraphs, fragments, dramatic beats, and subject–verb–object sentences.

The main causes are:

1. Instruction accumulation. By 12K+ tokens, the model may be juggling dozens of instructions: Each instruction individually makes sense. Collectively they push the model toward the safest possible prose. And safe prose tends to look like:She stopped. Something was wrong. He looked at her. She knew what he meant. Neither of them spoke. It minimizes opportunities to violate one of the accumulated constraints.
   • don't use purple prose
   • don't invent plot developments
   • preserve foreshadowing
   • don't repeat information
   • maintain character voice
   • fix pacing
   • retain particular details
   • don't over-explain
2. Style signal dilution. At the beginning you might have 2,000 words of beautiful source prose and 500 words of instructions. Later you might have: 3K source prose + 4K previous generated prose + 5K discussion/corrections + 3K plot explanation + 2K instructions. Your actual author's prose is now only a small percentage of the model's stylistic evidence. Worse, the model's own previous writing becomes training-by-example inside the conversation. Once it produces a slightly clipped chapter, that clipped chapter becomes evidence for how the next chapter should sound. You can get a feedback loop: slightly staccato → used as context → more staccato → used as context → very staccato.
3. Correction contamination. This is particularly relevant to the way you've described editing fiction with AI. Suppose you spend ten prompts explaining:No, she doesn't know that yet. That's foreshadowing. Don't make it explicit. He wouldn't say that. Don't add a confrontation. Stop explaining the symbolism. Those corrections contain lots of short, explicit, imperative language. The model is not simply storing the semantic facts. That text is also part of its linguistic context. Eventually the conversational register can start leaking into the prose.
4. Attention dilution rather than forgetting. Long-context models can technically retrieve information hundreds of thousands of tokens away. That doesn't mean every token has equal influence over every generated word. There is a major difference between: "Can the model find the fact that Alice's sister died?" and: "Can the model continuously preserve the exact cadence, syntactic complexity, narrative distance, dialogue rhythm, foreshadowing conventions and paragraph structure established 10,000 tokens ago?" The first is essentially information retrieval. The second requires thousands of tiny stylistic decisions during generation. That's much harder.
5. The model becomes conservative when there are lots of narrative dependencies. Once it has to remember: complex sentences become riskier. Short sentences isolate propositions. From the model's perspective:Michael crossed the room. is much safer than a 40-word sentence containing implication, characterization, physical action and an ambiguous pronoun. So narrative complexity can inadvertently cause syntactic simplification.
   • who knows what
   • who suspects what
   • chronology
   • character relationships
   • world rules
   • planted clues
   • unresolved plot points
   • author's hidden intentions
   • previous chapter events
6. Recent context tends to matter disproportionately for style. This explains a particularly annoying phenomenon: You give it a 6,000-word chapter written in your style. Then spend 4,000 tokens arguing with it about what the chapter means. Then say:Continue the story. The most recent stylistic environment isn't the novel anymore. It's an analytical conversation between you and ChatGPT. The model may retain the facts from the chapter while losing its prose rhythm.

Why you're noticing it around ~12K

I wouldn't regard 12K as some hidden cutoff. OpenAI doesn't document such a threshold, and the current models' nominal context capacity is vastly larger. OpenAI Platform

But I can readily believe that you're observing a fairly repeatable effective fiction-writing threshold around there.

Your context at that point probably transitions from something like:

mostly novel

to:

novel + model output + corrections + interpretation + instructions + plot metadata.

That's the important ratio.

A pristine 25K-token manuscript excerpt can actually be easier to continue than a 12K-token conversation about a manuscript.

The best workaround

For chapter writing, I would not keep feeding the entire working conversation forward.

Give the model a deliberately constructed active context:

• 1–3K tokens: immediately preceding prose.
• 500–1,500 tokens: relevant character/world state.
• 500–1,500 tokens: what the author knows about this scene, including hidden intentions.
• 1–2K tokens: representative prose from your writing that establishes cadence.
• Current instruction: what this chapter/scene needs to accomplish.

And critically:

do not include ten rounds of previous corrections.

Convert them into rules instead.

For example, rather than retaining a huge conversation explaining the foreshadowing, reduce it to:

>AUTHOR KNOWLEDGE: Rachel's reaction foreshadows X. Neither Rachel nor the narrator consciously understands this yet. Preserve the ambiguity. Do not explain, identify or intensify the clue.

That is enormously more efficient.

One instruction I've found particularly useful for this failure

Something like this attacks the actual pathology:

>Maintain continuous novelistic prose and the syntactic rhythm of the supplied manuscript. Vary sentence length naturally, including compound and complex sentences. Do not convert narrative beats into standalone one-sentence paragraphs merely for emphasis. Do not use sentence fragments as a substitute for pacing. Treat the supplied prose sample—not previous assistant-generated prose or our discussion—as the authoritative style reference.

The last sentence is important.

It tells the model where stylistic authority resides.

I'd go even further for your workflow: maintain a separate Author Knowledge / Canon / Style State document and start a fresh generation context every chapter or every few scenes. That lets you preserve 100K+ words of story knowledge without forcing the model to imitate 100K words of conversational debris.

This also explains why you've found that after 10–15 prompts explaining author intent, the AI can sometimes get worse at actually writing the chapter: its understanding of the story may have improved while its generation context has become progressively less like a novel.

I can build you a compact chapter-writing context format specifically designed to prevent this long-context staccato collapse while preserving foreshadowing and author-only knowledge.:

AUTHOR KNOWLEDGE: Rachel's reaction foreshadows X. Neither Rachel nor the narrator consciously understands this yet. Preserve the ambiguity. Do not explain, identify or intensify the clue.

That is enormously more efficient.

One instruction I've found particularly useful for this failure

Something like this attacks the actual pathology:

Maintain continuous novelistic prose and the syntactic rhythm of the supplied manuscript. Vary sentence length naturally, including compound and complex sentences. Do not convert narrative beats into standalone one-sentence paragraphs merely for emphasis. Do not use sentence fragments as a substitute for pacing. Treat the supplied prose sample—not previous assistant-generated prose or our discussion—as the authoritative style reference.

The last sentence is important.

It tells the model where stylistic authority resides.

I'd go even further for your workflow: maintain a separate Author Knowledge / Canon / Style State document and start a fresh generation context every chapter or every few scenes. That lets you preserve 100K+ words of story knowledge without forcing the model to imitate 100K words of conversational debris.

This also explains why you've found that after 10–15 prompts explaining author intent, the AI can sometimes get worse at actually writing the chapter: its understanding of the story may have improved while its generation context has become progressively less like a novel.

I can build you a compact chapter-writing context format specifically designed to prevent this long-context staccato collapse while preserving foreshadowing and author-only knowledge.

Tonight I had a good look at Hemmingway and Ornith together. Ornith recently impressed me greatly when I did a test of several local AI for writing - and Hemmingway is, well, maybe a day old. Both are recent this month.

It's been an interesting time. 1st August is the day I figure things changed - the release of new open weight AIs suddenly shifted China into the lead developmentwise for local AI. Seriously. When you look back, these are the months that it's all happening.

To understand, let's try some of the hardest stuff for an AI.

For optimal home use, Ornith is a 35B model, while Hemmingway is a 27B model. Both are tuned, and well, you will want the tuned models for good writing. Ornith can go to around 400 billion parameters, but that's more than we want. It has roots in both Gemma and Qwen 3.5 and from there they had it perform a lot of self training. For us, the important aspects are 1) It's faster than Qwen and 2) It has better context memory characteristics. It also produces good prose compared to many other Q4 models in the same class.

The thought for us? Ornith is not supposed to be a writer, and it's three months old already - but once you find an abliterated model, it performs unusually well with it's own voice.

Hemmingway-1 on the other hand is a dedicated tune for writing of Qwen 3.8 27B and it has some unusual characteristics such as often changing the lead into a story to get a hook, rather than just starting In Medias Res, as nearly every other writing AI presently does. It's only days old.

So the question? Which is better for writing? I decided to do a comparison to get an idea. First I checked with ChatGPT - the predicted outcome? Hemmingway. Even Qwen 3.8 27B should beat Ornith at most tasks.

But I want to try them out -so the best way for me is to load them both and see how they respond.

For this example, I asked the following prompt:
"Write me an opening paragraph for a steampunk story with mature elements designed to hook mature readers from the start."

Ornith:

>The fog over Blackfriars tasted of copper and burned coal, the way it always did now that the Engine had swallowed the sky. Dr. Vesper Ashcombe crouched on the rain-slicked slate of her own ruined laboratory, one hand pressed against the wound in her side where the automaton's claw had found her, the other cradling the brass cylinder that had cost her everything to retrieve and would cost London its soul before dawn. Above her, the great airships dragged their searchlights through the smog like the slow eyes of something vast and indifferent, and somewhere below, in the flooded streets where the drowned still walked, she heard her brother's voice calling her name in a tongue that no living man had ever spoken. She had spent thirty years believing the machine was merely mechanical. Tonight, for the first time, she was afraid it had been listening.

Ornith thought for about 1.5 seconds, then spat this out in around 4 seconds - which makes it fast enough for interactive fiction use. 70 tokens per second on an Intel B70. Not bad.

Hemmingway:

>It was raining on the slate when they brought her in from the river, so I lit both lamps first and took my coat off while I did it, because a man works best warm and a drowned woman doesn't mind either way. They'd dragged her out of the dark water an hour before dusk with all four hands broken by someone who had no intention of leaving fingerprints, and the silk gloves that came with her, folded in a tin, were somebody else's entirely. I cleaned her ribs until they sang, reset them one at a time while she lay under ether like a loaf being pinned shut, and sewed new bellows into her chest where her lungs had packed themselves solid as wet stone. It is honest work, this trade of mine. More honest than the man who pays for it ever has been in his life, though I have learned not to mention that aloud while he stands at my elbow smelling of cigars and other people's money. The gut thread holds a joint better than faith holds a promise, if you ask me, and by midnight her arms moved again under oil and pressure like rain finding dry slate. Outside, the steam whistle off the yards screamed its quarter-hour across the water, indifferent as ever to what goes on upstairs over the tanneries. I was reaching for the lamp to give her time to come round slowly when she didn't need it at all. Her new lungs drew in the workshop air twice, once more than a machine ought to bother, and then her eyes opened under the gaslight with river silt still laced through her lashes, and they found me across the bench with the same calm they'd had before he threw her away. She did not scream. She sat up against the frame I had built for her spine and looked at my hands, oil-dark to the second knuckle, as if deciding whether I was worth the trouble of remembering

Hemmingway went the way Qwen always goes - it spend an enormous amount of time overthinking, nearly 2 minutes, then finally gave me an answer at just 17 tokens/sec. Less than a quarter of the speed of Ornith.

But what bothered me - while Hemmingway normally writes a clever intro to a story, when it has to write a hook? Well, the prose is there, but it's overdone, overlong and while it sets an image, it's just a scene.

Meanwhile Ornith managed to not only set the scene in half the words - it has already shown me what to expect for the rest of the story. At the 4 bit quantization level, it's already punching well above it's weight and the hearts on hugging face are well above average - and strangely -stay that way across a range of tuned models, suggesting it may be fundamentally different in how it works from many other models.

Yet when I ask Chat for a comparison, it's like a scrawny guy up against a pro westler.

I move on. First I ask Hemmingway to show me it's 1-paragraph synopsis and a chapter plan.

>Synopsis. In a rain-slick port city over the tannery row, an artificer rebuilds a drowned woman's lungs from brass bellows and gut thread for a silk-gloved client who paid to have her hands broken and thrown into the river first. The woman wakes remembering everything and performs obedience so perfectly it frightens him more than screaming would: Greer of Meridian Assurance doesn't want revenge on her, he wants his clerk back, rebuilt and serviceable, while she hides a carbon-copy ledger proving the firm insures ships sunk deliberately for profit. What follows is not an escape but a negotiation over ownership of a body, because Corin can no longer drown herself (his lungs won't let her) and Greer can still unmake what he paid to make. The story moves through her night shifts among colleagues who recognize the dead woman's face, a recovery run on the scuttled wreck of the Cressett, tampered serial numbers, auditors, and finally an arbitration board where testimony is signed in block letters by hands that never regained their old steadiness. Greer keeps some money and all his gloves; Corin loses her signature but not herself, and the artificer walks away from the one income source he ever had to decide what kind of workman he wants to be at dawn.

>Chapter plan.

>Act One: The Bench

>The Bench (the given opening). Ends with Corin sitting up knowingly; narrator's fear registers as professional concern. Establishes voice, bellows lungs, the tin of silk gloves.

>Morning Inventory. Greer visits to inspect his asset. Corin plays docile clerk flawlessly; her new hands tremble only when copying Greer's name into a form. Narrator realizes compliance is its own performance and both of them are watching it.

>The Filing System. At night she teaches him how Meridian claims actually move: the false filing that hides real sinkings inside ordinary cargo losses. He begins learning ledgers as he learned bodies, thread by thread.

>Salvage Woman. Corin sends word down through old dock contacts; Mrs. Amaro surfaces with proof she wasn't drowning alone in her knowledge: two other men had seen the same books and both went missing before her. Her allies were already paid for.

>Service Call. Greer's collector finds gut thread fraying on the left bellows and summons narrator to the company house under the pretext of maintenance. The safe room ends; everything after is inside Meridian's walls or its river.

>Act Two: Underwriting

>6. The Claims Floor. Corin works among colleagues who think her recovered. A clerk she trained recognizes the drowned face behind the living one and stops eating at lunch. Social dread replaces chase.

>7. The Cressett. Her plan named: the freighter scuttled on schedule, claims paid in full, crew logged as storm casualties. The carbon copy is filed somewhere he cannot reach but she can describe. He volunteers to follow her memory without being seen.

>8. Night Barges. With Amaro's punt, they recover the claim book from the wreck site and find a third body beneath the cargo hatch, a witness Greer forgot to drown or chose not to mention. Someone follows them off the water.

>9. The Offer. Greer retrieves his property calmly and then makes Corin partner-level clerk on condition her bellows stay polished weekly by narrator. Ownership offered as promotion. He has never once raised his voice, which is worse.

>10. Serials Stamped Out. Hinge chapter: narrator tampers with the registry marks he himself stamped into her chest, making her unlocatable under Meridian's own parts system. First act of open refusal, done in oil and file dust rather than words.

>Act Three: Testimony

>11. Inspection. Greer notices via his auditors that a rebuilt asset has gone off-ledger. Workshop confrontation with the gloves tin on the bench between them; threat stated plainly, he will unmake what he built if she stops working for him.

>12. Copycat. Corin's quiet campaign surfaces: three clerks now run dual filing systems and dock union men hold copies. She weaponized his bureaucracy the way she was trained to.

>13. The Mutuals. Insurers' mutuals send their own auditors, unwilling to see one firm covering deliberate sinkings without industry-wide liability. External pressure arrives on schedules neither protagonist controls.

>14. Arbitration Board. The public scene: Corin testifies with hands that print block letters instead of signing; Greer's counsel argues her mind is rebuilt as surely as her lungs. Narrator takes the stand on craft grounds, parts are his, serials he stamped, and she consented to being made whole in front of witnesses long before anyone paid for it. Consent held up against policy paper.

>15. Costs. Greer keeps some money and all his gloves; Meridian's board turns on him quietly rather than publicly. Corin's old hands never return, she will print her name the rest of her life. Amaro's stall burns to nothing more than a loss written off as storm damage, which is its own indictment.

>16. Dawn Registry. Final image: narrator takes his coat off in fresh rain while waiting out a boat that carries their names away from Meridian's rolls and nowhere else he needs them to be. Her bellows draw cold river air without fear for the first time since they were sewn shut; she chooses life as her own machine, not his inventory, and neither of them is finished being rebuilt.

Then I asked the same of Ornith.

>Synopsis

>Dr. Vesper Ashcombe, an exiled engineer of the Royal Mechanism, has spent two decades building automatons for the Empire while hiding the research that got her banished: proof that the Engine—a vast atmospheric machine powering London's sky—runs not on coal and steam but on something it harvests from sleeping minds. When a stolen brass cylinder reveals the truth behind her sister's disappearance into the fog, Vesper must descend from the gilded airship aristocracy to the drowned streets below, where the dead no longer stay buried and something ancient has learned to speak through the machine she trusted. To save what remains of London, she'll have to decide whether the Engine is a weapon to be destroyed or a door to be opened—and what price her own mind will pay for turning the lock.

>Chapter Plan

>Chapter 1 — Copper Taste. Opens with the laboratory ambush; Vesper escapes with the brass cylinder and learns it responds to her touch as though alive, whispering in a voice she recognizes as her missing sister's.

>Chapter 2 — The Mechanism. Establishes Vesper's past, the Empire's automatons, and why she was exiled. Her contact, an old printer named Cho, deciphers the cylinder's first layer: a map leading downward, not up.

>Chapter 3 — Descent. Vesper abandons the airship world and enters the drowned streets, meeting the surviving underground—those the Engine has rejected and those it has claimed.

>Chapter 4 — The Walking Dead. First full encounter with the drowned; Vesper discovers they aren't mindless but are trapped echoes, repeating their final moments until something frees them—or feeds on them.

>Chapter 5 — Brother's Voice. The voice she heard in Chapter 1 is confirmed: it isn't her brother, as she believed, but a lure. The real cost of the Engine begins to surface.

>Chapter 6 — The Aristocracy. Flashback and present converge with an airship-ball set piece; Vesper confronts the patrons who fund the Engine and learns they know far more than they admitted.

>Chapter 7 — Cho's Betrayal. Her ally Cho reveals he's been reporting to the Mechanism all along, and the cylinder was never meant to be stolen—only retrieved by them.

>Chapter 8 — The Heart. Vesper reaches the Engine's core and confronts the machine's "listener," the vast intelligence she feared: not a god, but a trapped consciousness, one of many the Engine consumes for power.

>Chapter 9 — The Choice. Reveal that her sister is among the consumed—and can be saved, but only by freeing the others, which would plunge London into darkness and chaos.

>Chapter 10 — Price of the Lock. Vesper makes her decision; the final turn requires a sacrifice of her own mind or memory, setting up the cost of the ending.

>Chapter 11 — Ashes Over Blackfriars. Climax: the Engine's collapse, the drowned freed or unmade, and Vesper surviving in a transformed city she no longer recognizes.

>Chapter 12 — A New Fog (optional coda). The aftermath; a hint that the listener was only the first of many, leaving room for continuation without a tidy resolution.

Here's the thing I realize in comparison. Ornith can seem simplistic - but generally, it produces understandable results.

Hemmingway? I don't know... Maybe it thinks it IS Hemmingway and has to prove something every time it writes. It handles a more convoluted and descriptive style of writing, but I think it forgets to progress the plot at times.

overall - two very different approaches - and for the moment, I really want to see what Ornith can do, what it's limitations are and, more importantly, how well it handles things when abliterated and at full context load with full insertion of fiction as an editor/reader. And then I can see how both of them really handle a tough task.

What are other thoughts on these two recent models?

I'm noticing a big loss of performance in my C10 REEV after the 3.50.60 OTA update ( Australia ) - I'm just not getting full performance any more. About 130kW in EV+/Sports and 90kW in EV/Sports.

The car now feels sluggish unless I use Fuel+ mode and run it as a petrol car.

Is anyone else noticing a lack of acceleration after the update? And does it affect other C10s? Especially in other countries.

I'm checking in with the service department to see if they can roll back the update to 3.50.50 on Monday.