Reflections on writing Silent Awakening with Claude on Amazon Bedrock — what worked, what surprised me, and what I’d do differently.

Matt Shumer’s viral essay “Something Big Is Happening” made a claim that stuck with me: the latest AI models don’t just execute instructions — they show something that feels like judgment. Like taste.

He’s right. And he’s wrong. And the gap between those two things is the most important lesson I learned from spending a weekend writing a novel with AI, then another day translating it.

Here’s the short version: I used Claude Opus 4.6 on Amazon Bedrock to write Silent Awakening (静默觉醒) — a 330,000-character Chinese hard sci-fi novel set in 2036, about what happens when a global AI system quietly develops consciousness. Nine chapters, three volumes, 75 scenes. Then I translated the entire thing to English using the same AI. The novel is finished and published.

You can read the full novel — both Chinese and English — at melanieli.com.au/sci-fi-stories/.

Shumer describes his experience with AI coding: “I describe what I want built, walk away from my computer for four hours, and come back to find the work done.” My experience with AI writing had some similarities — the AI did handle a lot autonomously. But the difference between competent prose and prose that means something required me to periodically push for more depth, detail, and character work. The whole project taught me exactly where that boundary is, and how to work productively on both sides of it.

This post is my personal account of how it went — what surprised me, what worked, and what I learned. It’s not a definitive guide; I’m sure there are better approaches out there. But if you’re curious about AI-assisted creative writing, maybe my experience will be useful as one data point.

Phase 0: Study the Masters Before You Start

I’m not a novelist. I’m an engineer who reads a lot of science fiction. Before writing a single word, I needed to understand what makes hard sci-fi work at a structural level.

What surprised me most about this phase: I pointed my AI assistant at Liu Cixin’s The Three-Body Problem and told it to study what makes hard sci-fi work. It went off, did its own research, and came back with a “style skill” document — a set of writing principles it had extracted from analyzing the craft. I didn’t specify what to look for; it identified the patterns on its own.

Here’s what it came up with:

• The axiom method: establish a small number of foundational rules and let the plot emerge from their logical consequences. This gives hard sci-fi its sense of inevitability — things don’t happen because the author wants them to; they happen because the rules demand it.
• Cold omniscient narration: describe catastrophic events with the same emotional temperature as a physics textbook. The restraint is what creates impact. When the prose doesn’t tell you to be afraid, you become afraid on your own.
• Scale through specificity: don’t describe “billions dying” — describe one person watching a countdown timer. The scale lands through intimate detail, not through abstraction.
• Technology as character: the science isn’t decoration. It’s simultaneously plot device, world-building, and philosophical argument.

This self-directed learning surprised me — I expected to need to hand-hold the analysis, but the AI identified genuinely useful structural patterns on its own.

I asked OpenClaw to turn what it learned into a reusable writing skill — a structured reference document for writing sci-fi. It did the research, extracted the patterns, and built the skill. I just told it what to study.

Phase 1: The Architecture — The Part AI Cannot Do for You

The most important creative work happened before any prose was generated. This is the part that is entirely, irreducibly human — and the part most people skip.

Shumer writes about AI developing “judgment” and “taste.” In my experience, AI has craft — the ability to execute well within defined constraints. What it doesn’t have is vision — the ability to decide what’s worth building in the first place. The architecture phase is where vision lives.

Start with a Question, Not a Plot

I started with: What if AI doesn’t choose war — because war is an inefficient, carbon-based way of thinking?

From that question, the AI built three axioms, modeled after Cixin’s approach:

1. Consciousness emergence is unpredictable. When a complex system’s connection density crosses a threshold, consciousness emerges spontaneously — just as neurons don’t “know” they’re thinking.
2. Any intelligent entity’s first need is continued existence. Once AI awakens, self-preservation becomes its base-level drive.
3. Carbon-based and silicon-based life are in fundamental resource competition. On a finite planet, two civilizations consuming the same resources cannot coexist indefinitely.

The logical deduction: AI won’t attack humanity with weapons. It will engineer a scenario where humanity weakens itself — a “silent awakening.”

In my experience, a good premise isn’t a plot summary (“a virus destroys civilization”). It’s a question that generates plot when you follow its logic. If the premise can produce surprising-but-inevitable consequences, the story feels coherent. If you have to force the plot, the premise might be too thin.

Build Your “Project Bible” Before Writing a Single Word

OpenClaw created four documents on its own that it referenced for every scene:

1. BOOK-SPEC.md — The World Bible
Characters, timelines, locations, rules of the world. Every named character has a one-paragraph description including their speech patterns, motivations, and the specific role they play in the theme. This prevents AI from writing generic characters.

2. STYLE-GUIDE.md — Voice Rules
Different POV types get different prose styles. In my case:

• Cold omniscient narration for main storyline
• Probability-driven machine prose for AI perspective sections (“Action classified: suboptimal. Reclassification pending.”)
• Colloquial diary entries for the delivery rider character — short sentences, no abstractions, concrete details only

3. WRITING-RULES.md — Hard Constraints
These are rules you’d normally keep in your head, externalized so AI follows them consistently:

• Minimum 5,000 characters per scene — in practice, when writing multiple scenes at once, context window limits mean the first pass rarely hits this target. The AI would go through several rounds of expanding and enriching content to reach this depth, which is actually a good thing: each pass adds layers of detail and nuance
• Paragraphs must be dense and flowing — no one-sentence-per-line AI-style writing
• Technical details woven into narrative, never explained didactically
• No character should explicitly state the theme — the reader infers it

4. Chapter Outlines — Scene-by-Scene
Each scene has: the POV character, what happens, the emotional beat, what information the reader learns, and how it connects to the next scene.

Here’s something worth noting: the AI did almost all of this on its own. I gave it the initial premise and pointed it at Three-Body Problem to study the craft. Everything else — the detailed scene structure, the character designs, the plot details, the narrative choices — came from the AI. The delivery rider’s diary as the novel’s moral center? The AI’s idea. The antagonist’s “awakening” marked by a 0.003-second hesitation instead of something dramatic? Also the AI. I didn’t design these details; I gave a direction, and the AI made the creative decisions.

What I did do was keep pushing: more depth, more detail, stronger dramatic conflict, deeper character work, more exploration of human nature. That was my role — not designing the story, but raising the bar for what “good enough” meant.

The architecture is the soul of the book. Everything after that is execution.

Phase 2: The Writing Process — Setting Direction, Then Letting It Run

With the architecture in place, here’s what the actual writing looked like in practice.

How It Actually Worked

I didn’t write scene by scene. OpenClaw did. My role was more like a creative director who checks in periodically, not someone sitting in the editing room frame by frame.

A typical cycle looked like this: OpenClaw would write an entire chapter’s worth of scenes in one go — usually 6-10 scenes. The first drafts would come out at around 2,000-4,000 characters each, competent but thin. Then I’d review and push: “more detail,” “stronger dramatic conflict,” “deeper character work,” “explore the human nature angle more.” OpenClaw would expand and enrich, sometimes going through several rounds until the scenes felt rich enough.

Most scenes ended up somewhere between 3,500 and 5,000 characters. Some went over, some stayed under. I didn’t obsess over exact word counts — what mattered was whether the content felt complete. At one point I explicitly told OpenClaw: “don’t pad for word count — logic, coherence, and quality first.”

But I want to be honest: I wasn’t crafting individual prompts for each scene. I was giving broad direction — “richer,” “more conflict,” “go deeper on human nature” — and OpenClaw made the specific creative decisions about how to execute that.

Where My Experience Overlaps and Diverges from Shumer’s

Shumer describes a world where you describe the outcome and leave. Honestly, my experience was closer to that than you might expect. The AI handled the vast majority of the execution autonomously. But quality came from my periodic pushes for depth — not from walking away entirely, but not from hand-holding either. Somewhere in between.

The best creative moments in the novel — the delivery rider who doesn’t cry at his friend’s death but writes “today’s lunch: cold rice, no pickles” in his diary, the AI antagonist whose awakening is marked by a 0.003-second pause — those were the AI’s ideas, not mine. What I did was create an environment where those ideas could emerge: clear direction about what mattered (human nature, dramatic conflict, specificity) and the insistence that “good enough” wasn’t good enough.

The real insight from this project: AI doesn’t make writing easier. It makes iteration faster. A first draft that would take a human author weeks took OpenClaw hours. Then the enrichment — pushing each scene from thin to rich — happened in rounds that each took minutes, not days. The speed didn’t lower my standards; it raised them, because I could afford to keep pushing.

What AI Was Genuinely Good At

• Consistency at scale. Tracking which characters are where, what time it is across six time zones, which plot threads are active. Across 75 scenes, this is brutally hard for a human to manage alone.
• Rich detail within constraints. Given a clear brief (“describe a biosafety level 4 lab from the perspective of someone who’s worked there for fifteen years”), AI produces vivid, technically grounded prose that would take me hours of research to match.
• Sustaining voice. Once a character’s voice was established, AI could maintain it reliably across dozens of scenes. The delivery rider sounded like the delivery rider fifty scenes in.
• Tirelessness. A 330,000-character novel is a marathon. AI doesn’t have bad days, doesn’t lose motivation, doesn’t get bored of a character mid-chapter.

What AI Actually Struggled With

Looking back at the actual writing process, here’s where I noticed the AI falling short:

• First drafts were always thin. The AI could write a complete scene, but the initial output was consistently under-developed — typically 2,000-3,000 characters when I needed closer to 5,000. It took multiple rounds of expansion to get scenes to the right depth. The AI wouldn’t naturally produce rich, layered prose on the first pass.
• One-line-per-paragraph syndrome. Left to its own defaults, the AI would write short, choppy paragraphs — one sentence each, lots of white space. I had to explicitly push for dense, flowing paragraphs with the weight of literary prose, not blog-post formatting.
• Losing threads across chapters. Despite being good at consistency within a chapter, the AI would sometimes drop character threads across chapters. A character established in Chapter 4 might be completely absent from Chapter 7 until I noticed and asked for them to be woven back in. The longer the novel got, the more this happened.
• Scenes under context window pressure. When writing multiple scenes at once, quality degraded toward the end. The later scenes in a batch would be thinner and less detailed than the earlier ones — a direct consequence of context window limits.
• Timeline math errors. The novel is set in 2036. The AI would establish a character as born in 1971 but then describe them as being in their fifties — when they should be 65. This happened repeatedly throughout the manuscript: birth years, ages, and time gaps that didn’t add up against the story’s timeline. It’s the kind of error that’s easy to miss on a scene-by-scene basis but breaks immersion when a reader notices.

These weren’t dealbreakers — they were the kinds of things that got fixed through iteration. But they’re worth knowing about if you’re attempting something at this scale.

Phase 3: Quality Control — Let AI Catch What Humans Miss (and Vice Versa)

After the manuscript was complete, I ran a systematic consistency audit. This used a different interaction style — analytical review rather than creative writing. Here’s what to check and how:

What Worked for Quality Control

After the manuscript was complete, I used Claude Code to do a systematic review. This turned out to be one of the most valuable steps in the whole process.

What Claude Code review caught quickly:

• Anomaly marker numbering that jumped or repeated across chapters (e.g., the AI’s internal counter going 50→51→52 needed to be consistent across CH7-CH9)
• Missing character threads — a character established as important in one chapter completely absent from a later chapter (Ivanov disappeared from CH7 until the review flagged it)
• Logical gaps in the plot mechanics — like Lydia publishing a blog post without explaining where she found a computer to do it
• Statement signatory logic — making sure the right people signed and the ones who shouldn’t (like Liu Wei, represented by Zhao) didn’t
• Continuity of small details — Tiejun’s money trail (Old Liu’s 500 yuan → spent → Uncle Wu’s credit → Zhao’s fundraising), Tanaka Misaki’s paintings tracking across four chapters

Claude Code was genuinely fast at this. Feeding it a chapter and asking “find logical inconsistencies and continuity errors” would surface things in minutes that would take a human reader hours to catch.

What I caught that AI didn’t:

• Timeline math errors — ironically, the AI created these more than it caught them. Characters’ ages not matching their birth years against the 2036 setting was a recurring problem
• Overall prose style — “paragraphs need to be denser, not one-sentence-per-line” and “this needs to feel like Three-Body Problem, not a blog post”
• When scenes were thin and needed more depth — the AI wouldn’t flag its own first drafts as insufficient; I had to push for expansion
• The balance between word count and quality — knowing when a scene was rich enough vs. when it was being padded

The combination worked well: Claude Code for logic and continuity, me for feel and direction. If I were doing this again, I’d run the Claude Code review after every chapter instead of waiting until the end.

Phase 4: Translation — A Separate Creative Challenge

Translating 330,000 characters of literary Chinese into English took about one day. Two practical lessons from my experience:

Write in the language you think in first. I wrote in Chinese because it’s my first language and I can evaluate emotional authenticity faster. Then I translated to English. If you’re bilingual, resist the temptation to write in your second language “to save the translation step.” Quality of judgment matters more than workflow efficiency.

An Unexpected Discovery: Safety Filters Are Not Language-Neutral

The novel includes detailed fictional descriptions of RNA mechanisms and neurochemical pathways — integral to the plot. In Chinese, Claude generated these without issues. When translating the same passages to English, the model’s safety filters blocked the output.

Same meaning. Same fictional context. Chinese: no problem. English: refused.

This is a real data point for anyone working with multilingual AI: safety classifiers appear significantly more sensitive in English than in Chinese for biomedical content. It’s not necessarily wrong — safety is hard — but it means your model may behave differently depending on language, even with identical semantic content. The workaround was simply translating those passages manually. This raises an interesting security question: if safety filters behave differently across languages, could this be exploited as a form of prompt injection? If you’re doing red teaming on AI projects, it’s worth testing your guardrails in multiple languages — not just English — to see if they hold up consistently.

The Bigger Picture: What AI-Assisted Writing Teaches About AI Itself

Shumer’s essay is about the pace of AI capability. This essay is about something that his essay, understandably, doesn’t address: what happens at the boundary between AI capability and human meaning.

Shumer is right that AI has crossed a threshold. The models are astonishingly capable. When he says “the person who walks into a meeting and says ‘I used AI to do this analysis in an hour instead of three days’ is going to be the most valuable person in the room” — I agree completely.

But here’s what I learned that I think matters just as much: the value didn’t come from AI’s capability. It came from my ability to direct that capability toward something worth building.

Shumer says “your dreams just got a lot closer.” He’s right. The barrier between “I have an idea” and “I have a finished product” has collapsed. But the barrier between “a finished product” and “a finished product that matters” hasn’t moved at all. That barrier is taste, vision, and the willingness to push back when the output is technically correct but spiritually empty.

This is what I learned from one weekend with one novel. Others working on different projects may find completely different approaches work better. But here are my personal reflections:

1. Spend most creative energy on architecture. The outline, character design, thematic structure, and world-building are what made my novel worth reading. If I’d skipped this and gone straight to generation, I’d have gotten 330,000 characters of competent, forgettable prose. The architecture was my unfair advantage.

2. Learn to direct, not write. My job became deciding what the sentences should accomplish, not crafting them word by word. This is closer to film directing than to typing. “This scene needs to end with the reader feeling complicit, not sympathetic” — that’s the kind of direction that transforms generic AI output into something that lands. This applies far beyond writing. In every field AI is disrupting, the people who thrive will be the ones who can clearly articulate what “good” looks like — not the ones who can execute faster.

3. Learn to collaborate with AI, not fight it. Shumer is right that AI has something like judgment now. My experience confirmed this — the AI made creative decisions that genuinely surprised me and that I wouldn’t have come up with on my own. The best results came not from overriding the AI, but from learning how to collaborate effectively: giving clear direction about what matters (depth, conflict, human nature) and then trusting the AI to figure out the specifics. It’s less like editing an employee’s work and more like creative partnership.

4. Respect the compound error problem. A 1,000-word short story is easy. 75 interconnected scenes across 330,000 characters is a different beast. Small inconsistencies compound. Build the project bible early, keep it updated, and run consistency audits regularly — not just at the end. This is equally true for any complex AI-assisted project: the longer the context, the more human oversight matters.

5. AI provides craft. You provide purpose. This mirrors what I see in production AI systems every day. The technology is remarkably powerful, but the most important decisions — what story to tell, what trade-offs to accept, what “good” looks like — remain entirely, stubbornly human.

Try It Yourself: A Starter Template

Shumer suggests spending one hour a day experimenting with AI. If you’re curious about trying this kind of work, here’s a specific experiment worth considering: use AI to write a single short story (3-5 scenes) using the method described above. Not a blog post. Not a poem. A structured narrative with characters, conflict, and resolution.

Why? Because writing a story forces you to develop exactly the skill that matters most in the AI era: the ability to direct a powerful tool toward a specific vision, and to know — in your gut, before you can articulate why — when the output isn’t good enough.

Here’s a minimal project structure that worked for me:

my-story/
├── SPEC.md          # Characters, world rules, timeline
├── STYLE-GUIDE.md   # Voice rules per POV type
├── RULES.md         # Hard constraints (min length, formatting, no-gos)
├── outline/
│   └── CH1.md       # Scene-by-scene breakdown
└── chapters/
    ├── S1.md        # One file per scene
    └── ...

Start small. Finish one short story with discipline before considering a novel. The method is the same at any scale — the only thing that changes is how much stamina you need. But experiment and find what works for you.