After my Claude Sonnet 5 vs Fable 5 vs Opus 4.8 post 'I tested Claude Sonnet 5 vs Fable 5 vs Opus 4.8 for fiction writing' a few people asked how I got some of the outputs to sound so good, especially the dragon scene.

I think the honest answer is: it was not one magic prompt. The thing that helped most was giving the model a harder story problem before it started writing. The easy version of that dragon prompt is pretty obvious:

• The crown lied
• The dragon is innocent
• The knight realizes this
• Mercy wins

That can still produce a nice story, but it is a very easy one. cos the model by default gets to solve the story by revealing that one side is good and the other side is bad.

The better outputs kept the pressure alive:

What if the dragon really did burn villages? What if saving the child restarts a civil war? What if doing the merciful thing gets someone else killed What if the knight’s oath is corrupt, but breaking it still costs something real?

That, more than “make it epic” or “write in a literary style,” was what changed the output.

I think this is where a lot of AI fiction prompting goes wrong. We ask for better output|story|prose when the real issue is that the model is solving an easy story.

If the underlying story problem is too clean, better prose just makes a clean story prettier. If the problem has real stakes, even imperfect prose has something to work with.

The setup I used tried to do four things:

1. Tell the model what kind of genre pressure I wanted: not just “epic fantasy,” but mythic scale in the background, concrete sensory detail in the foreground, and a choice that costs something.
2. Warn it away from the obvious shortcut: for this prompt, the shortcut was “dragon good, crown bad.” Once the dragon is fully innocent, the moral dilemma mostly disappears.
3. Keep the scene embodied: the knight should be seeing, touching, avoiding, arguing, noticing. If the dragon just explains the whole plot in a monologue, the scene gets weaker.
4. Make the choice happen on the page: A lot of AI fiction ends with “he had until dawn to decide.” That can sound good, but often it means the model avoided the actual dramatic action.

So my takeaway is that “better AI writing” is not always about asking for prettier sentences. A lot of the time, it is about shaping the problem the model is trying to solve.

If anyone wants to try the same kind of setup, the free epic fantasy generator is here: Free story generator

It uses the hand-crafted epic fantasy profile I was testing with. The profile itself is not exposed or downloadable, but you can try the behavior for free and compare what it does against your own manual prompts.

Disclosure: I work on Noren, so obviously I’m biased. But I’m sharing this because people asked what was happening under the hood, and I think the underlying method is portable even if you never use our tool.

The queen was dying while the western watchfires went out one by one.

From the high chamber, Seluin could see them through the rain: nine red points on the black shoulder of the mountains, then seven, then four. Each lost fire meant a tower taken or abandoned. Each tower meant families running east with what they could carry, soldiers falling back through mud, granaries left open to the winter wolves and the men who had learned from them.

Inside, the queen’s breath scraped like a saw through wet wood.

“Not the princes,” she said.

No one answered her at first. The chamber was full of people who had trained all their lives to answer queens. Tonight they were trying to decide which silence was least treasonous.

Prince Halec stood at the foot of the bed in armor bright enough to shame the candles. His younger brother, Tovan, wore black silk and a face made for mourning. Behind them waited captains, priests, sworn judges, the lord of coin, and three envoys from the river cities whose ships had not yet chosen a side. The realm had come to fit inside one room, and the room smelled of fever, rain, old wool, and fear.

Seluin stood near the brazier with a tray of cooling broth in her hands.

She had been called because the queen still took water from no one else. For twelve years she had served in the palace kitchens, then in the sickrooms, then at the queen’s side when the wasting began. She knew how much honey to stir into bitter medicine. She knew which linen did not scratch. She knew the sound the queen made when pain had become too large for pride.

She did not know why the queen was looking at her.

“Come here,” Queen Oratha said.

Seluin set the tray down. Every eye followed her across the tiles. She felt the old burn of being watched by those who knew her place and saw her stepping out of it.

“Majesty,” she said, kneeling.

The queen’s hand moved beneath the blanket. On her thumb was the ruling ring, iron dark around a ruby like banked fire. It had been forged, people said, from the nail that sealed the first king’s coffin after he bargained with the mountain powers. Seluin had polished it during fevers. It had always seemed too heavy for any living hand.

Oratha lifted it with effort.

“The crown will take ash before dawn,” the queen said. “The western oath has broken. The thing under Kaul has woken, and my sons will spend the realm arguing who may command the army.”

What do you think?

I’m trying something different from cold email. We're building a writing app, and I’m looking for bloggers who actually cover writing tools, AI tools, blogging workflows, or author software.

Not looking for a guaranteed positive mention, and I’m not dropping a link here. I’m more interested in finding people who would be open to testing something and giving a real opinion, even if that opinion is critical.

If you write about this kind of thing, how do you usually decide what is worth reviewing? Do you prefer founder emails, free access, a short demo, or just a clear product page?

Also happy to read a few blogs from people here if you cover this space.

It is amazing what changes when AI knows how you build and present ideas.

>A prompt tells the model what to do, a profile tells it who is doing the writing.

Remember this every time you write with AI

I ran a small voice-preservation test and the result will surprise you

The setup:

1.  Same generation model for both outputs: Claude Sonnet 4.6
2.  Same prompt
3.  Same format: short social/journal post
4.  Same source writer
5.  The only thing that changed was the extracted voice profile

>One profile was extracted by Claude. The other was extracted by GPT-5.5/Codex. Then both profiles were fed into the same Claude generation route.

Prompt: Write a short-form social/journal post about starting a new job, doing paycheck math after taxes, insurance, retirement, rent, groceries, and gas, realizing more money still does not mean freedom, and deciding to get serious about budgeting. Do not mention this is a test. Do not use markdown.

Output using the Claude-extracted profile:

So the new job is official, ya'll, and for about five minutes I genuinely felt like I had finally gotten ahead of something. Then I sat down with my calculator and did the actual math.

After taxes, health, dental, and vision, the life insurance I'm paying for just in case, and the retirement contribution I know better than to skip, the gap between what I thought I was making and what actually hits my account is genuinely disrespectful.

Rent comes first. Groceries. Gas, which has opinions about my finances that I did not ask for. All the random things I keep forgetting to factor in until they're already in my cart.

*insert the slow, humbling realization that making more money and being financially free are not the same thing*

I've been here before, making more, spending more, somehow ending up staring at the same account balance wondering where it all went. The gross number on the paystub is not the number that matters, and I know this, I've always known this, but knowing something in theory and actually building a life around it are two completely different things.

That's just it. More money without a plan is still just money that leaves faster. So I'm done doing this casually, done with the version of budgeting where I feel vaguely optimistic at the start of the month and confused by the end of it. Every dollar in and every dollar out, tracked for real, not the aspirational version where I round down on groceries and pretend I don't eat out that much, with a real budget and numbers I don't look away from.

I can't wait to find out what my life actually costs. LOL. *shrug*

Output using the Codex 5.5-extracted profile:

So I started the new gig this week and spent most of orientation just thinking about how much better my life was about to get, which, in retrospect, was optimistic.

Then I ran a paycheck calculator.

Between federal and state taxes, health insurance, and the retirement fund I signed up for because I'm genuinely trying to be responsible about this, I'm already looking at close to $250 gone before I even see a check. Add rent on top of that, then groceries, then gas because I drive further now, and all in all...I'm making significantly more than I was and somehow still not free. Just broke at a higher income bracket.

That's a specific kind of demoralizing I was not prepared for.

I refuse to keep doing this, so I pulled out my budget spreadsheet and I'm treating it like a requirement now, not something I look at when I feel like it. I know I'm wasteful and there's really no excuse for it even if I understand why, and that changes now.

More money doesn't automatically mean breathing room. It just means your mistakes cost more.

Full disclosure: I’m building Noren, so this is part of my own testing around voice profiles. But I’m not linking it here because I’m more interested in the question:

Which output feels more like it preserved a real writing voice, and why?

I’m building Noren, a writing tool for people who use AI for customer-facing business writing but hate how generic the drafts sound.

The problem I kept seeing is this:

AI can help with emails, how-to replies, follow-ups, FAQs, proposals, and customer updates. But the draft often comes out sounding too polished, too neutral, or too much like a template.

For a small business, that matters. Customers are usually not interacting with a faceless brand. They are interacting with the owner, founder, or team directly. If the writing suddenly sounds like Joe instead of you, it can hurt trust.

Noren works by creating a reusable voice profile from your own writing samples. Then you can use that profile to make AI drafts sound closer to how you actually communicate.

It’s meant for things like:

  - customer emails

  - support replies

  - follow-ups

  - how-to explanations

  - newsletters

  - proposals

  - social posts

I’m not trying to replace the human pass. The goal is to make the first draft start closer to your voice, so you spend less time rewriting generic AI output.

If anyone here uses AI for customer-facing writing, I’d be curious where it breaks for you: emails, support, sales follow-ups, proposals, or something else?

Link: https://usenoren.ai

I’m asking partly because I’m building around this problem, so full disclosure.

I’ve been thinking a lot about AI-written customer emails. They’re useful for getting a first draft out quickly, especially when you’re tired or trying to explain something clearly. But the drafts often sound a little too polished or generic. Sometimes they’re technically fine, but you read them and think: this does not really sound like me.

For a small business, that feels risky because customers are often dealing with the owner directly. If the email suddenly sounds like a template, it can change the trust a bit.

Curious how other owners handle this.

Do you let AI draft customer emails? Do you rewrite heavily? Or do you avoid using it for anything customer-facing?

We launched Noren today on PH. Appreciate your supports.

What is Noren?

Noren turns your writing samples into a portable voice profile, so AI can write in your voice across emails, posts, newsletters, fiction, and other long-form drafts. Not a tone slider, a profile built from your real writing.

If you like this, check us out on PH.

Thanks