<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Mathieu Demarne on Medium]]></title>
        <description><![CDATA[Stories by Mathieu Demarne on Medium]]></description>
        <link>https://medium.com/@mdemarne?source=rss-12a25aa39744------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*a9gG6qbsdh6pnKGvvjKykg.jpeg</url>
            <title>Stories by Mathieu Demarne on Medium</title>
            <link>https://medium.com/@mdemarne?source=rss-12a25aa39744------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 03:20:21 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@mdemarne/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Writing with AI: What Works, What Doesn’t, and What’s Left]]></title>
            <link>https://medium.com/@mdemarne/writing-with-ai-what-works-what-doesnt-and-what-s-left-72d53d95a6da?source=rss-12a25aa39744------2</link>
            <guid isPermaLink="false">https://medium.com/p/72d53d95a6da</guid>
            <category><![CDATA[writing-tools]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Mathieu Demarne]]></dc:creator>
            <pubDate>Thu, 12 Mar 2026 01:12:22 GMT</pubDate>
            <atom:updated>2026-03-14T20:49:45.217Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IDP2V4MOZ6sdC9-kclfJTw.png" /><figcaption>Source: Image generated by the author \ Disclaimer: this story is edited with the help of Author Kit (AI).</figcaption></figure><p>In the <a href="https://medium.com/@mdemarne/catching-what-drifts-in-your-human-led-ai-assisted-manuscript-4d4fb7334a24">previous article</a> on my experiment writing with Author Kit, I ended with a question: <em>does all this governance produce better writing, or only more consistent writing?</em></p><p>I want to answer that honestly. The workflow I have described across this series (Chapters: <a href="https://medium.com/@mdemarne/why-ai-writes-the-same-book-every-time-0b02323f618a">1</a>, <a href="https://medium.com/@mdemarne/the-world-your-ai-forgot-to-build-c7fda0fe71c7">2</a>, <a href="https://medium.com/@mdemarne/catching-what-drifts-in-your-human-led-ai-assisted-manuscript-4d4fb7334a24">3</a>) prevents several predictable failures. It does not magically produce genius.</p><p><strong>The question is whether that tradeoff is worth it for the kind of project you are writing.</strong></p><p>This chapter is a snapshot of where models and tools stand in early 2026, not a timeless verdict. They will keep evolving, and some implementation details will date.</p><p>Let me start with the cost.</p><h3>The Real Cost</h3><p>Structure takes time. Governance takes attention. Every artifact I have described, the constitution, the concept, the world files, the outline, the chapter plans, the review passes, they all require work before a single word of prose appears.</p><p>The constitution alone can take an hour or more to write well (even if Author Kit can help you). Often, I tend to put excerpts of prose with the style I like and would like to see replicated, and finding the right tone is a quest of its own. The concept file, with clarification rounds, can consume an afternoon. World building scales with complexity: a short story needs almost nothing, but a multi-perspective novel with intertwining timelines requires a serious investment in entity tracking. That’s up to the user to decide. The plan, draft, review loop for each chapter is slower than open-ended generation, sometimes significantly.</p><p>If you are writing a three-thousand-word essay, this is probably overkill. Unless, like me, you are not a native English speaker or you would rather have the LLM be a checking your idea structure, and style.</p><p>If you are writing two hundred thousand words with interdependent rules, recurring characters, and an argument that builds across chapters, the calculus changes. The question is not whether structure costs time. The question is whether structure costs less than the alternative. The alternative is rework, drift, and the slow accumulation of local decisions that contradict each other.</p><p>Another cost besides time is the cost of the models themselves.</p><h3>What Worked</h3><p>The constitution did what I hoped it would do. It constrained voice.</p><p>Without a constitution, every generation pass draws on the full statistical distribution of AI training data. The prose drifts toward the average of what the model has seen. With a constitution, the model has a narrower target. First person present, short paragraphs, no hedging, no marketing language. The voice I defined became the voice that showed up on the page.</p><p>This was not automatic. The first draft of any chapter still needed review and often constitutional amendments. But the distance between first draft and final was smaller than it would have been without those constraints. I was revising execution, not fighting drift. And there I could see and measure progress.</p><p><strong>The world and outline artifacts prevented the most obvious coherence failures.</strong> In the cave story I used as an illustration in past articles, cars duplicated because nothing said they could not. In this series, nothing like that happened. Because the structure made that kind of failure visible in advance.</p><p>The analyze and revise loop caught issues before they compounded. In article two, I made a claim about fact consistency versus structural consistency. In article three, I almost introduced a subtle shift that would have contradicted the framing. The analysis pass caught it. The revision was minor. The downstream impact would have been visible.</p><p>The reconciliation step, checking upstream documents against drafted prose, saved me from planning future chapters against an outline that no longer matched reality. That alone is worth the overhead for long projects.</p><h3>What Didn’t Work</h3><p>Some problems remain unsolved.</p><p><strong>Analysis quality depends on rule quality. The constitution can enforce voice, but it cannot manufacture novelty.</strong> If I define generic rules, I get generic compliance. The tooling amplifies intent. It does not substitute for it. In many ways, using Author Kit frees the author from the actual writing and gives room to focus on the novelty. Substance over style in a world where both are needed. And the author as full control to jump in at any time.</p><p><strong>Humor is particularly hard.</strong> I can write a constitution that says <em>“dry humor permitted, no slapstick,”</em> but the AI still struggles with timing. A joke that works on the third attempt is not really a joke. It is a slot machine that eventually paid out. It still does better than me, however, as I am not very good at making jokes.</p><p><strong>Surprise has the same problem. A well-structured outline prevents contradictions, but it also reduces the chance of happy accidents.</strong> The model follows the plan. The plan was mine. If the plan was unimaginative, the prose is unimaginative with better continuity. The inspiration of the writer that materializes while writing is lost a bit.</p><p><strong>Some outputs remain technically consistent but emotionally flat.</strong> The sentences are correct. The structure holds. The reader does not care. I cannot trace this to a specific failure in the workflow: it feels more like a ceiling on what current models can do when guided rather than led. Guiding a model toward a tone is different from giving it a reason to <em>feel</em> that tone.</p><p><strong>Temporality is particularly hard.</strong> I mean here the handling of time, timelines, and temporal logic in narratives. If a story involves flashbacks, parallel timelines, or time manipulation, the model struggles even with explicit rules. The cave story was an extreme case: duplicates emerging on consecutive days, each copy with a different timeline of experiences. One chapter would assume events that were not yet described. Working hard on the constitution and planning eventually fixed this in most cases, but it took time. The world file helped track state. The outline helped track obligations.</p><p><strong>And then there is the iteration problem.</strong> Nothing works out the first time. I ran multiple revision cycles on every article. The process is faster than starting from scratch each time, but it is not instant. Anyone who tells you AI-assisted writing eliminates revision has not shipped a manuscript (or code).</p><h3>What Surprised Me</h3><blockquote>Most of what improved the final output was decided before I wrote a single word of prose.</blockquote><p>That felt counterintuitive at first. I expected the generation step to be where the magic happened. Better prompts, better models, better samples. The reality was different. The constitution and concept did more to shape the result than anything I did during drafting. The world and outline did more to prevent failure than any fix I applied after.</p><p>Drafting became less magical but more reliable. I knew what I was going to get before I generated it. The surprises were smaller, more local, more manageable. That tradeoff, less surprise for more control, is not obvious. It took me several projects to accept it: but it works.</p><p>The upstream work is also where the human contribution matters most. <strong>The AI can draft. It cannot decide what the book should be, who it is for, what voice it should have, what counts as good.</strong> Those decisions live in the constitution, the concept, the outline. The model executes. The human directs. The division is not philosophical. It is practical. If you skip the direction, the execution defaults to average.</p><h3>Cross-Model Reality Check</h3><p>Different models carry different default cadence and stylistic priors. That is not speculation, it is often very observable.</p><p>The first article in this series was written with Claude Opus 4.6. The second with GPT 5.2. The third with GPT 5.3 Codex. This one with GLM 5.</p><p>You likely could not tell. Not because the models are identical, but because the constitution and style anchor treated voice as canonical and generation as variable. Each model produced prose that passed the same review against the same standards. The differences in default tone, sentence rhythm, and metaphor density were normalized during revision.</p><p>That is the point of the workflow. It does not eliminate model variation but it constrains it to a target that the human defined. The result is consistent output even when the underlying model changes.</p><p>Names and numbers follow the same principle. Without intervention, models gravitate toward defaults. They pick similar names, similar quantities, similar patterns. Author Kit handles this by generating candidates once (with strong prompt <em>against</em> using averages), selecting deliberately, and forcing lookup reuse so the manuscript does not reinvent people or rediscover a favorite number every time the model changes (ask for a number and hope for something else than 7 or 42). <strong>The workflow reduces default gravity. It does not create true randomness.</strong> That remains a limitation.</p><p>Mixing high-tier models with low-tier models is not something the workflow fixes. If you draft with a frontier model and revise with a smaller one, the quality gap shows. Prompts cannot equalize raw capability. What the workflow can do is keep the voice consistent across capability tiers, so the reader does not feel the handoff.</p><h3>The Overhead Question</h3><p><em>Is this just waterfall methodology for writers?</em></p><p>I understand the criticism. A constitution, a concept file, a world database, an outline, chapter plans, review passes. That sounds like the kind of process-heavy approach that agile software development spent twenty years dismantling.</p><p>The distinction matters. Waterfall front-loads all decisions and locks them. Author Kit front-loads intent and keeps decisions revisable. The constitution defines voice. It does not prevent me from changing my mind about voice. It makes me do it explicitly. The outline defines structure. It does not prevent me from pivoting. It makes me see the impact before I commit.</p><h3>Practical Endpoints</h3><p>Author Kit includes export paths. Word (docx) for traditional publishing workflows. Epub for ebook distribution. An audiobook path through text-to-speech integration (for now using OpenAI):</p><pre>authorkit book audio --voice onyx<br>authorkit book build --format epub</pre><p>I am mentioning them here for completeness: they are delivery mechanics, not quality mechanics. The workflow ends when the manuscript is ready, not when it is formatted. But those allow me to generate new chapters and listen to them while driving and correct them after.</p><h3>Portable Principles</h3><p>If you never install Author Kit, the series still has value as the principles transfer:</p><ul><li><strong>Define intent before generation.</strong> A constitution or style guide, written before you draft, constrains the output space. It does not matter whether you use a tool or a text file. What matters is that you make your expectations explicit before the AI tries to infer them.</li><li><strong>Separate generation from verification.</strong> The model produces text. You verify against intent. Do not merge these steps. If the tool that generates also judges, you lose the ability to catch drift.</li><li><strong>Track changes at the entity level.</strong> Characters, places, rules, claims. Anything that can drift should have a record. The record does not need to be fancy. It needs to be external to the final output and linked to chapter references.</li><li><strong>Keep human decisions explicit.</strong> When you change direction, write down what changed and why. When you park a decision, record it with a deadline. When you reconsider a fact, trace the impact. The workflow does not need to be automated. It needs to be visible.</li><li><strong>Accept that revision is the work.</strong> AI-assisted writing is not faster if you skip the part where you make the result good. The speed gain comes from not losing coherence across long projects, not from eliminating the editorial pass.</li><li><strong>Know your threshold.</strong> Not every project needs this much structure. Not every writer works this way. The value is proportional to complexity and interdependency. For short pieces, the overhead is waste. For long manuscripts, it is leverage.</li></ul><h3>Conclusion</h3><p><em>So: better writing or only more consistent writing?</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IkGfGySvbW7pYJBZPu7twA.png" /><figcaption>Source: Image generated by the author</figcaption></figure><p>The honest answer is that structure improves outcomes, but it does not bring brilliance. What it guarantees is that the manuscript you ship is the one you intended to write (if you phrased that intent and its context explicitly). The voice holds. The logic holds. The argument or story does not drift into contradiction with itself. The characters do not randomly change eye color or motivation. The rules do not mutate in chapter seven because the model inferred a pattern you never authorized.</p><p><strong>Whether that manuscript is brilliant still depends on you. </strong>The AI cannot supply originality, genuine surprise, or emotional resonance on demand. It can execute against your definition of those things. It cannot define them.</p><p>Every tool in the workflow defers to you at the decision point. The constitution is yours. The pivot requires your approval. The what-if branch waits for your verdict. That is not a limitation of the tooling. That is the premise.</p><p>You can adopt these principles without installing a single tool. You can also use Author Kit, if the workflow fits your process. The discipline transfers either way.</p><p>In these articles, I tried to answer this question:</p><blockquote><em>How do you write coherently with AI without producing work that reads like everyone else’s AI output?</em></blockquote><p>The answer is not a prompt. The answer is a process. Define intent, build structure, verify output, iterate. The model does the generation. You do the directing.</p><p><strong>The directing is still the hard part.</strong></p><h3>Trying Author Kit</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/875/0*y6dikcHdUxPfxwOt.png" /><figcaption>Source: Image generated by the author</figcaption></figure><p><a href="https://github.com/mazemerize/author-kit">Author Kit</a> is open source. However, you will need an AI Coding agent at your disposal (Claude Code, Codex or GitHub Copilot) and a Python environment with uv. I acknowledge that for many writers, this can be quite tech oriented. Later, if any interest, a better UI/UX could be developed into a product.</p><p>Install Author Kit by running:</p><pre>uv tool install authorkit-cli --from git+https://github.com/mazemerize/author-kit.git</pre><h3>References</h3><p>You can find this series of articles and the blueprint of how they were generated by Author Kit <a href="https://github.com/mazemerize/gen-author-kit-articles"><strong>here</strong></a>.</p><p>You can find the reference story, <em>The Cave at Low Tide</em>, <a href="https://github.com/mazemerize/gen-cave-at-low-tide"><strong>here</strong></a><strong>.</strong></p><ul><li>It serves as a demonstration of Author Kit’s ability to<strong> generate coherent, internally consistent prose</strong> over a long narrative.</li><li>During the writing process, Author Kit was given <strong>broad latitude to shape the plot</strong> and develop character arcs. As such, this book is<strong> not </strong>intended to showcase what it would produce in close collaboration with an author and does not stand as a testament that AI can write good prose, but rather to illustrate how Author Kit supports structural and narrative consistency.</li><li>Multiple models were used interchangeably throughout the project, including Claude Opus 4.6, Sonnet 4.5, GPT-5.2, and GPT-5.3 Codex.</li><li>You can listen to the generated audio of the first chapter <a href="https://github.com/mazemerize/gen-cave-at-low-tide/blob/main/book/dist/audio/01-chapter-1-the-camping-trip-that-wasnt.mp3"><strong>here</strong></a>.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=72d53d95a6da" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Catching What Drifts in Your Human-Led, AI-Assisted Manuscript]]></title>
            <link>https://medium.com/@mdemarne/catching-what-drifts-in-your-human-led-ai-assisted-manuscript-4d4fb7334a24?source=rss-12a25aa39744------2</link>
            <guid isPermaLink="false">https://medium.com/p/4d4fb7334a24</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[writing-tools]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[software-development]]></category>
            <dc:creator><![CDATA[Mathieu Demarne]]></dc:creator>
            <pubDate>Thu, 05 Mar 2026 15:20:03 GMT</pubDate>
            <atom:updated>2026-03-05T15:38:01.162Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*pxRYFNOx4Btn5O-z" /><figcaption>“Mosaic Tile” by George Hodan \ Disclaimer: this story is edited with the help of Author Kit (AI).</figcaption></figure><p><em>This article continues a series about writing with AI, and a framework I built to help with the process. You can read those first articles using the links below.</em></p><p><a href="https://medium.com/@mdemarne/why-ai-writes-the-same-book-every-time-0b02323f618a">In the first article</a>, I describe the failure mode that made me build Author Kit: unstructured AI generation drifts toward plausible averages until the story collapses. I also define the intent layer, constitution plus concept, so voice and goals are explicit before any prose gets generated.</p><p><a href="https://medium.com/@mdemarne/the-world-your-ai-forgot-to-build-c7fda0fe71c7">In the second</a>, I turn intent into structure: a world state that can be tracked, an outline treated as a contract, and a plan, draft, review loop that forces discipline before drift compounds.</p><p>In that cave manuscript I mention in the first two articles (<em>a person enters a cave at low tide and comes out twice, two days in a row</em>), the second and third chapters still read clean. The collapse only appears when you line the chapters up and notice the seam.</p><p><strong>That is the mosaic problem. Each tile looks fine. The image is wrong.</strong></p><p>This is where chapter level discipline stops and verification starts. Author Kit treats verification as a separate layer because drift is rarely a single dramatic mistake. It is a quiet accumulation of small ones.</p><h3>The Mosaic Problem</h3><blockquote><strong><em>The worst kind of inconsistency is the one you do not notice.</em></strong></blockquote><p>If a chapter is obviously broken, you fix it. If it is subtly wrong, you keep building on it. You turn a local mistake into a global direction. In a series of essays, drift is not usually a timeline contradiction. It is softer and more dangerous.</p><p>The claim changes a shade between articles. In my first article I talk about models as “averaging machines”. In this one I start talking as if they were only unreliable assistants. The reader cannot tell you where the flip happens. They only feel the series become less coherent.</p><p>Or the voice shifts.</p><p>Nothing in the argument is technically wrong, but the cadence changes. A few paragraphs get longer. The metaphors get more ornate. The sentences start sounding like a helpful assistant instead of a person with skin in the game. It happens fast if I get lazy and stop enforcing my own constraints.</p><p>Or the thread simply disappears.</p><p>In the first article I promise that this is not about prompt tricks. In the second I show structure. If this one forgets to explain how structure is actually verified, the series loses its spine. The reader remembers the promise even if they do not remember the sentence.</p><p>That is why I treat this as a verification article. The goal is not to add more structure. The goal is to catch drift that structure alone cannot prevent.</p><h3>Analysis as Read Only Verification (Prose CI)</h3><p>In software, I do not merge code because it compiles on my machine. I merge it because checks run, and because those checks are biased toward finding the failures that show up late.</p><p>In the first article I compare this approach to spec-driven development (SDD): write the spec first, generate the artifact second, and then verify the artifact against the spec. For this series, the constitution, concept, and outline are the spec. The drafts are the artifact. Analysis is the CI pass that keeps them aligned.</p><p>That is what Author Kit does for prose with /authorkit.analyze.</p><p>It is a read only pass. It does not rewrite chapters. It produces a report with findings, severity, and suggested next actions.</p><p>This matters because of how these projects fail. If a tool quietly edits text while it is “checking” it, I cannot trust the check. I cannot tell whether the output is good, or whether it has simply been coerced into looking consistent.</p><p>Author Kit keeps analysis and mutation separate on purpose.</p><p>The analysis pass also stays high signal. It caps itself. Fifty findings is the limit by design, not a performance constraint. If a report can produce two hundred nits, it will. And if it does, I will ignore it.</p><p>Severity is what makes the report usable. When analysis finds a constitution break, voice slip, tense break, nonnegotiable style violation, the framework treats it as CRITICAL and stop to fix it first. When reader visible coherence is at risk, a claim is contradicted elsewhere, a prerequisite appears too late, a promised thread vanishes, that is HIGH. MEDIUM means the series still works but drift has started: cadence flattening, loose logic, references that need tightening. LOW is polish, useful later but not worth derailing active drafting. This hierarchy allows to focus on what matters most first and avoid modifying everything all at once.</p><p>In this series, imagine article one framing models as averaging machines, then a later piece quietly reframing them as mere assistants with no bridge between the two. No single sentence is obviously wrong. The series still loses coherence. That is why analysis runs seven lenses every time: <em>continuity, argument coherence, themes and motifs, pacing, voice and style, thread setup and payoff, and governance around parked decisions and deadlines.</em></p><p>Author Kit renders the result as a table:</p><pre>| ID | Category | Severity | Location(s) | Summary | Recommendation |<br>|----|----------|----------|-------------|---------|----------------|<br>| A1 | Voice | MEDIUM | Article 2 | Cadence shifts across the series | Revise to match the constitution |<br>| A2 | Continuity | HIGH | Article 1, Article 3 | Article 3 assumes X before Article 1 explains it | Add a bridge or reorder |</pre><p>The point is that with this, drift becomes visible enough to manage.</p><h3>Reconciliation (When the Specs Drift Too)</h3><p>There is another drift that is more embarrassing, because it is not the model’s fault: <em>it is when “</em><strong><em>my documents lie”.</em></strong></p><p>I do not mean maliciously. I mean in the way every project lies after its first revision. I change a paragraph because it reads better. I adjust an ending because it sets up the next article more cleanly. Then my outline still describes the older version. My plan still describes the older version. A note I wrote during planning still implies a claim I removed.</p><p>If I keep planning from stale documents, the series becomes inconsistent even if every chapter is well written.</p><p>Author Kit treats drafted chapters as canonical for that reason.</p><p>It can scan for factual claims made by upstream documents about already drafted chapters, then verifies them against the prose that actually exists. When there is a mismatch, the prose <em>can</em> wins. The outline and plans are updated after — and this is the important part.</p><h3>Fixing Drift (Revise, Pivot, Retcon)</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1rr-oWk80tPjMSzl2zOMhQ.png" /><figcaption>Tacoma Narrows Bridge twisting and buckling in high winds shortly before its 1940 collapse.</figcaption></figure><p>Once analysis finds drift, the next question is not what to edit. The next question is what kind of change I am making.</p><p>Most writing workflows blur this and create chaos. Everything becomes “edit the chapter.” That works until the first real direction change.</p><p>Author Kit keeps three actions separate because the risks are different.</p><ul><li>/authorkit.revise is what I use when direction is still correct and execution is not. A paragraph drifts in tone. A transition gets muddy. A claim is correct but poorly staged. Revision is targeted surgery with local risk.</li><li>/authorkit.pivot is different. I am changing direction on purpose. Maybe this series should center model switching earlier (this actually happened!), or spend less space on tool taxonomy and more on governance. That is not a bug fix. It is a scope decision with cascade risk across concept, outline, and chapter plans.</li><li>/authorkit.retcon is the sharpest tool. An established fact changes retroactively, and every downstream consumer has to move with it. In prose, that is the same risk profile as refactoring a public API: partial application is worse than doing nothing, because nine updates and one miss create a contradiction that looks intentional.</li></ul><p>The workflow looks simple when the category is clear.</p><pre>/authorkit.analyze<br>/authorkit.revise Address the top 3 CRITICAL findings<br>/authorkit.retcon &quot;duplicates only copy humans&quot; -&gt; &quot;duplicates copy humans and their clothes&quot;</pre><p>Those commands make the type of change explicit so the rest of the workflow can behave correctly.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/888/1*e2kghffU9VxtWMkV9qtzNg.png" /><figcaption>Example of retcon report from “The Cave at Low Tide”. Generated by the Author.</figcaption></figure><h3>Safe Experimentation (Snapshots and What If)</h3><p>Not every decision is a fix. Some are exploration.</p><p>In writing, especially with AI in the loop, the only honest way to evaluate a direction is to draft it. The problem is that drafting commits you. You now have prose that either becomes canon or becomes sunk cost.</p><p>Author Kit treats experimentation as a workflow primitive using /authorkit.whatif.</p><p>Snapshots are bookmarks with context. It uses the local folder git infrastructure (Author Kit requires — or assumes — you are using a version control system). A snapshot is not only a git tag. It also records why I bookmarked the state, what is working, and what I am uncertain about. That context matters later when I compare outcomes.</p><p>What if branches are controlled experiments. Author Kit creates an experimental branch, takes a snapshot first, and records a hypothesis. Then I can try the alternate direction without polluting mainline.</p><p>The key difference is that comparison is narrative, not only textual. A diff tells me what lines changed. A narrative comparison tells me what the reader experience changed.</p><p>That is what makes experimentation safe. It lowers the cost of admitting uncertainty.</p><h3>Decision Parking (Momentum Without Amnesia)</h3><p>Some uncertainty is not worth exploring immediately.</p><p>The failure mode is familiar. I notice a question. I keep writing. I forget to answer it. Then I reach the end and discover that three chapters depend on a choice that I never made.</p><p>Decision parking is Author Kit’s way of preventing that: /authorkit.park.</p><p>When a decision is parked, it is written down with a deadline, urgency, and options considered. The point is not to solve it. The point is to make it impossible to accidentally forget it. A decision may need to be made about Chapter 2, but it won’t matter until we start working on Chapter 12. We can wait until then.</p><p>This is also where analysis becomes more than continuity checking. If a parked decision is overdue, it should show up as a finding. Not because the tool is trying to force my hand, but because the workflow needs to surface risks before they become retroactive clean up.</p><p>There is a subtle benefit here too. Parking is a way to keep writing without lying to myself. I do not have to pretend the uncertainty does not exist. I do not have to resolve it prematurely. I can keep momentum and keep honesty.</p><h3>Changing the Rules of Duplications in the Cave Story</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*wchxc0NX1kMfjve6amYnjg.png" /><figcaption>Photo by Jeremy Mosley</figcaption></figure><p>This is where the cave story becomes useful again, because it contains the exact kind of change that breaks long projects.</p><p>Suppose I write half the manuscript with one duplication rule, and then I realize the rule is wrong.</p><p>Maybe duplicates should age faster. Maybe the duplication only works at the spring equinox, not any low tide. Maybe the cave duplicates clothing after all, because otherwise the story keeps tripping over naked copies and the joke gets old.</p><p>This is not a continuity error. It is a creative decision.</p><p>But it is also a cascade.</p><p>The moment the rule changes, every downstream reference to the rule becomes suspect. Any paragraph that relied on the old rule must be updated, including the indirect consequences. If duplicates age faster, it changes alibis, motivations, and what counts as urgent.</p><p>This is the kind of change that makes unstructured AI writing collapse, because the model has no global map of which chapters depend on which facts. It only sees the local context window, and the window is not a memory. It is a sampling of recent text.</p><p>Author Kit treats this as a sequence, not as a vibe.</p><ol><li><em>First, </em><em>/authorkit.snapshot. Lock the state so I can compare later if the change was a mistake.</em></li><li><em>Then, </em><em>authorkit.pivot. Generate an impact plan across artifacts. What does this rule touch in the concept, the outline, the world, and the drafted chapters.</em></li><li><em>Then, </em><em>/authorkit.retcon. Produce a manifest of direct references, indirect references, and derivative consequences. Show the list before editing anything.</em></li><li><em>Then, </em><em>/authorkit.revise the chapters that were affected, keeping the local voice intact.</em></li></ol><p><em>Then, </em><em>/authorkit.analyze again. Not because the tool is trying to win an argument, but because retroactive considerations are where humans miss the tenth reference.</em></p><p>The point of this workflow is not to make direction changes painless. Direction changes are not painless. The point is to make them tractable, so the cost is visible and the results are consistent.</p><h3>Conclusion</h3><p>This article is not a promise that Author Kit produces perfect prose. It is a claim that process help, especially with AI. But drift is inevitable. Models average. Humans forget. Documents rot.</p><p>The difference between a project that survives and a project that dissolves is whether the workflow catches drift before the reader does. Author Kit adds structure. It adds gates. It adds verification.</p><p>Does all this governance produce better writing, or only more consistent writing? The next article will try to answer.</p><h3>Trying Author Kit</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/875/0*shTUUkFhcG8ob4PP.png" /><figcaption>Source: Image generated by the author</figcaption></figure><p><a href="https://github.com/mazemerize/author-kit">Author Kit</a> is open source. However, you will need an AI Coding agent at your disposal (Claude Code, Codex or GitHub Copilot) and a Python environment with uv. I acknowledge that for many writers, this can be quite tech oriented. Later, if any interest, a better UI/UX could be developed into a product.</p><p>Install Author Kit by running:</p><pre>uv tool install authorkit-cli --from git+https://github.com/mazemerize/author-kit.git</pre><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4d4fb7334a24" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The World Your AI Forgot to Build]]></title>
            <link>https://medium.com/@mdemarne/the-world-your-ai-forgot-to-build-c7fda0fe71c7?source=rss-12a25aa39744------2</link>
            <guid isPermaLink="false">https://medium.com/p/c7fda0fe71c7</guid>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[writing-tools]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[software-development]]></category>
            <dc:creator><![CDATA[Mathieu Demarne]]></dc:creator>
            <pubDate>Thu, 26 Feb 2026 06:01:15 GMT</pubDate>
            <atom:updated>2026-03-02T02:06:22.701Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*N1j4f0SMJ5V-Yo9FuwxDsw.png" /><figcaption>Source: Image generated by the author \ Disclaimer: this story is edited with the help of Author Kit (AI).</figcaption></figure><p><a href="https://medium.com/p/0b02323f618a">In the first article</a>, I defined the core failure mode of generating complex stories with LLMs: unstructured AI generation drifts toward plausible averages, and the story collapses. I also lay down the first response layer, <em>constitution plus concept</em>, so intent is explicit before generation starts.</p><p>This article picks up from there and answers the next question: <em>how do I keep that intent coherent once the manuscript starts to grow?</em></p><p>Remember my cave story, which I use as a logic test for AI systems: A person enters a cave and comes out <em>twice</em> on two separate days. By chapter three, the model is duplicating cars, and nobody authorized that rule.</p><h3>The Breakdown Is Structural, Not Random</h3><p>That sentence still annoys me because it sounds absurd and avoidable at the same time. In my cave manuscript, the original premise is clear: one man enters a cave at low tide and gets duplicated. The first two chapters stay mostly on track. Then chapter three starts to leak logic. If humans can duplicate, maybe jackets can duplicate. Maybe keys can duplicate. Then cars duplicate. From the model perspective, that inference is plausible. From the story perspective, it makes no sense.</p><p>The failure is not dramatic. It is incremental. The gangster copy suddenly talks like the politician copy. A copy who is in London shows up in Buenos Aires with no transition. A relationship thread that should get sharper gets blurry. If you read any single chapter in isolation, it still sounds fluent. The collapse only appears when you read across chapters.</p><blockquote><strong>That is the part that makes unstructured AI writing deceptively risky. Local fluency disguises global incoherence.</strong></blockquote><p>The chapter in front of you reads cleanly. The manuscript behind it drifts. If you only evaluate sentence quality, you miss <em>system quality</em>.</p><p>This is the core thing to retain from the experience: large language models generate convincing local continuations. They do not maintain durable story state unless you give them a system that does.</p><p>The problem is the same, in fact, as with coding. I have learned it the hard way: give an AI a complex task without proper grounding and it will build things you did not ask it to. By the time you review its work, it is too late: you now have to understand it to fix it or throw it all away and start again. And overall, you would have been faster on your own.</p><p>Three controls are missing:</p><ul><li>No entity ledger that says what exists, how it changes, and what cannot change.</li><li>No chapter contract that says what this chapter must deliver, what thread it advances, and what it must not contradict.</li><li>No quality gate loop that checks each chapter against both the plan and the established world before it moves forward.</li></ul><p>Once the problem is framed this way, the solution gets less mysterious. Author Kit stops searching for better phrasing in prompts and starts building structure around generation.</p><blockquote><strong><em>Engineer the AI so it can do the engineering. Or, in this case, the writing.</em></strong></blockquote><h3>World Building Is an Entity Database</h3><p>When people hear “world building,” they often imagine static files: a map, some backstory, maybe a note about what people eat for breakfast in the northern provinces (or wherever else the story takes place). That can be fun. It is not enough for long form AI writing.</p><p>In practice, Author Kit treats world building as operational memory.</p><p>Each world file tracks an entity that matters to the manuscript. That entity can be a character, a place, a system, an organization, even a recurring object if the object has narrative obligations. The point is not completeness but traceability. If an entity appears in chapter two and changes in chapter six, Author Kit needs a record that ties those states together, with chapter references.</p><p>Each world entity splits into two field families. <strong>Identity fields</strong> answer what this thing is and what else it is called. <strong>State fields</strong> answer what is true <em>right now</em>. Identity should be stable. State is expected to evolve. If those are mixed carelessly, every edit looks like a rename and every rename looks like a plot change. Once that confusion enters a long manuscript, continuity debt piles up fast.</p><p>Character entities need role, current objective, and relationship deltas. Place entities need constraints and access logic, not just sensory description. System entities need explicit boundaries and exceptions. The common pattern is this: <em>do not store trivia first, but store whatever can break causality if it drifts.</em></p><p>That is why Author Kit uses a YAML frontmatter on world entries. Machine readable metadata makes retrieval and validation straightforward.</p><pre>---<br>id: SYS_DUPLICATION_RULE<br>aliases: [low tide rule, cave rule]<br>chapters: [CH01, CH02]<br>evolution: [(CONCEPT), (CH01), (CH03-rev)]<br>---</pre><p>A tiny header like that does a lot of work. It gives a stable identity, alias lookup, chapter scope, and a minimal change history. In prose terms, it answers four critical questions: <em>what</em> is this, <em>where</em> does it appear, <em>what else</em> calls it by another name, and <em>when</em> did its definition change.</p><p>Author Kit can then programmatically build a lookup index under world/_index.md. This is fast, as it does not require AI. This acts as a retrieval layer that lets planning and drafting tools resolve references without guessing. If chapter planning mentions &quot;the cave rule&quot; and the world entity is named SYS_DUPLICATION_RULE, alias lookup bridges that gap. If /authorkit.retcon (the command for retroactive edits) changes a rule later, the model can see where that rule is used.</p><p>Such an index is important, as semantic meaning is often not as clearly defined in free-form text as in code. It is easy to search where a method of an object class is used. But it is much harder to find where a world concept is used, if that concept can be named in different ways, and sometimes used implicitly.</p><p>In the cave manuscript, one sentence in a system entity would have prevented the car problem:</p><p><em>“Only living organisms duplicate. Objects, vehicles, and infrastructure do not.”</em></p><p>That sentence enables scope control. It is a boundary condition in plain language, tracked as a first-class artifact instead of hidden in my head.</p><p>Memory inside a prompt is fragile, but in structured artifacts, it becomes durable. This is also why, as part of conception, Author Kit can use /authorkit.clarify multiple times to explore unclear areas of the book concept and force the AI to ask for the explicit information it needs to avoid vague assumptions.</p><h3>Fact Consistency Is Not Structural Consistency</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pL5uDXwpsDpc6JEYRN3loA.jpeg" /><figcaption>Source: The Architect’s Dream by Thomas Cole</figcaption></figure><p>At this point I need to make a distinction that gets lost in tool comparisons.</p><p>Tools like <a href="https://sudowrite.com/">Sudowrite</a>, <a href="https://www.novelcrafter.com/">NovelCrafter</a>, <a href="https://novelai.net/">NovelAI</a>, and <a href="https://www.squibler.io/">Squibler</a> solve real problems. They help track details. They help keep names, traits, and setting facts stable. Several also add meaningful in-app structure for drafting. If your main pain point is attribute drift and local context quality, these tools are useful and often faster than custom workflows.</p><p>But fact consistency is not the same as structural consistency.</p><p>Fact consistency asks: <em>does this entity still have blue eyes in chapter twelve?</em></p><p>Structural consistency asks: <em>does this entity still have the same role, motivation, causal path, and thread obligations across twelve chapters, including changes that are intentional and traceable?</em></p><p>Both matter but they are not interchangeable.</p><p>The cave manuscript makes this obvious. I can perfectly track eye color, age, and place names while still losing the core logic of duplication scope, copy identity separation, and consequence flow. The manuscript can look stable at a detail layer and still fail at the architecture layer.</p><p>This is why I position Author Kit with a different category emphasis, not as the only structured tool in the space.</p><p>The Author Kit workflow is built around decisions over time: what gets defined before generation, what gets validated after generation, and how midstream changes propagate safely. In software terms, this is closer to a development workflow than a writing assistant panel.</p><p>If I need a one hour brainstorming burst, this workflow is overkill. If I am managing twenty chapters with interdependent rules, it becomes leverage.</p><h3>The Outline Is a Structural Contract</h3><p>The outline is where many writing projects pretend to be disciplined, then quietly become improvisation with extra steps.</p><p>I do not say that as criticism. I say it because I have done this for years. I write an outline, feel organized for an hour, then drift the moment drafting gets interesting. With AI in the loop, that drift gets amplified because the model gladly follows the newest local pattern and treats your old plan as optional context.</p><p>In Author Kit, the outline is treated as a contract.</p><p>But it is important to note that a contract does not mean rigidity. It means explicit obligations. For each chapter, Author Kit defines purpose, summary, key beats, thread duties, and callback duties. Purpose answers why the chapter exists at all, key beats constrain what must happen, and thread duties track what long arcs need movement. Callback duties prevent the model from reintroducing information as if it were new.</p><p>This is especially important in a modular article series where each piece should stand alone but still reward sequential reading. If article two re-explains article one from scratch, reader attention gets wasted. If article two assumes too much, new readers get lost. The outline helps hold that line.</p><p>There is another important rule: drafted chapters are canonical.</p><p>This sounds obvious until you hit revisions. Suppose the outline says chapter one ends with framing A, but the actual drafted chapter ends with framing B because B reads better and I keep it. If I continue planning from stale outline A, continuity drift starts immediately. So planning must reconcile against drafted text, not against old planning assumptions. The outline stays a contract, but it is a living contract that gets updated when reality changes.</p><p>Back to the cave case: if the chapter contract says “<em>track each copy identity and location explicitly at each chapter boundary</em>,” the spontaneous London to Buenos Aires teleport does not survive even one review pass.</p><p>Prompt quality helps, but contract clarity helps more.</p><h3>The Chapter Lifecycle Is the Quality Gate</h3><p>Once world and outline artifacts exist, they still do nothing by themselves. They become useful when they are wired into a repeated loop.</p><p>In Author Kit, the loop is plan, draft, review.</p><p>In practice, this is the sequence used in a chapter workflow:</p><pre>/authorkit.chapter.plan 2<br>/authorkit.chapter.draft 2<br>/authorkit.chapter.review 2</pre><p>Each command is a gate. Author Kit moves forward when the previous step passes, not just because text exists.</p><ul><li>/authorkit.chapter.plan is where Author Kit captures chapter intent before prose appears. In fiction, that means scenes and emotional beats. In non-fiction, that means sections, argument progression, setups, and payoffs. This is where opening and closing obligations are defined.</li><li>/authorkit.chapter.draft is where Author Kit generates prose with context loaded on purpose: constitution for style and voice, chapter plan for local structure, and relevant world and outline entries for continuity boundaries. The draft is not an open ended jam session. It is an execution pass against declared constraints.</li><li>/authorkit.chapter.review is where Author Kit tests output against those constraints: plan adherence, constitution compliance, continuity against prior chapters (in style and substance), and world consistency where applicable. Findings are then classified for revision.</li></ul><p>The last step sounds procedural, so here is what it looks like in plain questions:</p><ol><li><em>Does this section do what the plan says it should do?</em></li><li><em>Does the voice still sound like the intended authorial voice, or like generic assistant prose?</em></li><li><em>Did this chapter smuggle in a claim that contradicts chapter one?</em></li><li><em>Did it introduce a new rule without recording it anywhere?</em></li></ol><p>If any answer is wrong, the fix should target the source of the mismatch, not random sentence polish.</p><p>The key is iteration discipline. Author Kit supports multiple revision cycles, even if a limited few is usually enough to move from rough but coherent to publishable quality for this kind of chapter or article. To fast-track this process, the command /authorkit.chapter can run this loop automatically up to three times.</p><p>This loop also makes reordering manageable.</p><p>In long projects, chapters move. Sometimes one chapter should be split in two. Sometimes two should merge. When chapters are treated as indexed artifacts with cross references, moving them is painful but tractable. Without that structure, reordering is chaos because references become latent bugs in prose form.</p><p>This matters more than it sounds. Reordering is not a cosmetic operation. It changes argument timing, reveal point, and reader trust. If a callback arrives before the setup after a reorder, the prose may still read smoothly but the logic is inverted. Structured chapter metadata makes these errors visible enough to catch.</p><blockquote><strong><em>A lot of AI writing advice focuses on generation prompts. A better question is </em>process topology<em>: where are the gates?</em></strong></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hIJXZcNtmEEl9cykdwMYCg.png" /><figcaption>Source: Image (mermaid graph) generated by the author</figcaption></figure><p>What gets checked before promotion to the next stage? If there is no answer, the workflow is relying on memory and vibes under growing complexity.</p><h3>Where Chapter Level Control Stops</h3><p>There is a limit to what chapter level discipline can guarantee. A chapter can pass review and still contribute to manuscript drift. Each tile in the mosaic can look fine. The full image can still be wrong.</p><p>Maybe chapter five quietly shifts a theme from ambiguity to certainty. Chapter six follows that shift. Chapter seven doubles down. No single chapter has an obvious defect. By chapter nine the book is making a different argument than intended, and no one can point to one catastrophic moment where it broke.</p><p>This is where chapter loops end and cross chapter analysis begins.</p><p>The pattern is familiar from engineering. Unit tests can all pass while the integration still fails. Chapter reviews are similar. They are necessary local checks. They are not proof that the whole manuscript remains aligned.</p><p>In the next article, I get into what catches that kind of drift: multi pass analysis across continuity, timeline, character consistency, pacing, voice, and unresolved threads. I also get into reconciliation and course correction, because structure itself can drift from drafted reality and because creative direction changes mid process whether you plan for them or not.</p><p>Intent without structure fails fast. Structure without verification fails slowly. And slow failure is harder to notice, because each local piece keeps looking plausible until the full system no longer holds.</p><p>That is exactly the kind of failure I would want my writing tools to catch before readers ever see it.</p><h3>Trying Author Kit</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/875/0*y6dikcHdUxPfxwOt.png" /><figcaption>Source: Image generated by the author</figcaption></figure><p><a href="https://github.com/mazemerize/author-kit">Author Kit</a> is open source. However, you will need an AI Coding agent at your disposal (Claude Code, Codex or GitHub Copilot) and a Python environment with uv. I acknowledge that for many writers, this can be quite tech oriented. Later, if any interest, a better UI/UX could be developed into a product.</p><p>Install Author Kit by running:</p><pre>uv tool install authorkit-cli --from git+https://github.com/mazemerize/author-kit.git</pre><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c7fda0fe71c7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why AI Writes the Same Book Every Time]]></title>
            <link>https://medium.com/@mdemarne/why-ai-writes-the-same-book-every-time-0b02323f618a?source=rss-12a25aa39744------2</link>
            <guid isPermaLink="false">https://medium.com/p/0b02323f618a</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[writing]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[Mathieu Demarne]]></dc:creator>
            <pubDate>Fri, 20 Feb 2026 17:32:51 GMT</pubDate>
            <atom:updated>2026-03-02T02:06:51.198Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vXJ4a0tJLckEjEx2CP43IQ.png" /><figcaption>Source: Image generated by the author \ Disclaimer: this story is edited with the help of Author Kit (AI).</figcaption></figure><p>I had this idea for a novel, designed as a test for AI:</p><p><em>A man discovers a loophole in reality: if he enters a coastal cave at low tide, he walks out twice. He is duplicated. Two of him wake up on consecutive days, each believing he is the original.</em></p><p><em>He starts exploiting it. He lies to his wife. He sends copies out to live the lives he never had the nerve to try. One becomes a gangster in Vancouver, slowly rising through a criminal network that does not know there are many of him. Another enters local politics on a platform of radical transparency, which is ironic given the circumstances. A third attempts absurd Guinness World Records. Eventually the copies multiply beyond control, the world buckles under the weight of too many of the same person, and they all have to decide which one gets to keep living. There can be only one.</em></p><p>The tone aims for Douglas Adams territory without tipping into parody. Fifty thousand words, give or take. I thought it was a good premise, especially given the multiple timelines and logic required to maintain them, to test on AI.</p><p>I gave it to an LLM. I provided the premise, the tone, the character setup, and a chapter of backstory. I fed each completed chapter back as context for the next. The first two chapters were promising. The voice was close to what I wanted. The humor landed more often than not.</p><p>By chapter three, cars were duplicating alongside humans.</p><p>I did not ask for that. The duplication rules, which I never formally wrote down anywhere, had drifted. The model decided that if humans duplicate in the cave, maybe objects do too. It seemed like a reasonable inference from the AI’s perspective. It was not. Then it lost track of which copy was where. Two copies were supposed to be in London, but one showed up in Buenos Aires without transition or explanation. Character identities blurred. The gangster copy started talking like the politician copy. Dialogue that was supposed to be menacing came out bureaucratic. The story did not collapse in one dramatic moment. It dissolved, paragraph by plausible paragraph. Each chapter, read in isolation, <em>felt</em> fine. The accumulated result was incoherent.</p><p>I tried to fix it. I added summaries of the rules at the top of every prompt. I tried lower temperatures, higher temperatures, different models, different reasoning levels. I rephrased the premise. Nothing held for more than a few chapters. The fixes were local. The problem was structural.</p><p>The diagnosis, in hindsight, is straightforward. There was no entity database telling the AI what can and cannot be duplicated. No outline constraining which copies do what in which chapter. No review cycle catching the drift between generations. The AI was generating the most plausible next sequence of tokens given its context window. It was not maintaining a world. It had no tool for it.</p><p>I am a software engineer working and living between two countries, Seattle, WA (USA) and Switzerland, where I did my studies. A few years ago, while working at Microsoft, I partnered with research to build a Copilot that helps engineers work with AI. That work taught me, concretely, what large language models are good at and where they quietly break.</p><p>While engineering is my profession, I have been writing on-and-off as a hobby since I was a teenager, in French and then in English even if I am not a native speaker. I have carried stories around for years. When LLMs arrived with the ability to generate prose at speed, I thought I had found a tool to help me correct my grammar, suggest edits, brainstorm ideas and overall write better. The cave story was my first test: it was full of potential pitfalls and loopholes. It became the proof that something else was needed.</p><p>Not because the AI lacked capability. Because without structure, capability defaults to the average. And I knew enough about how these models work to understand that the fix would not come from better prompts. It would have to come from somewhere else entirely.</p><h3>Why This Happens</h3><p>My cave story is not an outlier. It is the default outcome of unstructured AI generation, and there is research that says so.</p><p>A team of researchers published a study in PNAS called “<a href="https://www.pnas.org/doi/10.1073/pnas.2504966122">Echoes in AI.</a>” They asked GPT-4 to produce a hundred versions of a short story inspired by Kafka. In fifty of those hundred generations, the policeman in the story gave the exact same instruction. When they ran the experiment with human writers, the overlap was negligible. Each person wrote something distinct. The AI wrote variations on a theme, and the theme was narrower than anyone expected.</p><p><a href="https://dl.acm.org/doi/10.1145/3635636.3656204">A separate study</a>, presented at ACM’s Creativity &amp; Cognition conference, found that participants using ChatGPT produced less semantically distinct ideas than participants using other creative tools. The gap was not small. And the researchers tried the obvious fixes. They raised the temperature parameter, which controls how much randomness the model injects into its word choices. They varied the prompts. They added explicit diversity instructions. None of it closed the gap. The homogenization was not a parameter problem: it was an architecture problem.</p><p>The mechanism is not mysterious as LLMs are trained on the statistical distribution of existing text. When you ask for a continuation, they generate the most probable one. Most probable means most common. Most common means most generic. This is why AI characters tend toward the same name patterns (try to search <a href="https://www.nytimes.com/2025/12/03/magazine/chatbot-writing-style.html?utm_source=www.aisidequest.com&amp;utm_medium=referral&amp;utm_campaign=which-model-is-best-for-writing-research-and-images">Elena/Elara Voss</a>), and why plots follow the Hero’s Journey whether the story calls for it or not. Why dialogue defaults to a certain cadence, descriptions to a certain density, and emotional beats to a certain predictable arc. Experienced readers can identify AI prose in a few paragraphs. Not because it is bad, but because it is always the same kind of average good.</p><p>There is a term for this in machine learning: mode collapse. The model generates the same patterns regardless of how you vary the input. Your third AI chapter sounds like your first. Your villain talks like your hero. Your comedy settles into the same rhythm as your tragedy.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mGbat3-zf18gl4SDszl7yA.png" /><figcaption>Interest trends for Elana Voss (blue) and Elara Voss (red). Some generic names that AI models tend to generate as they converge to the same behavior have become characters across multiple stories in the past few years. Source: <a href="https://trends.google.com/explore?q=Elena%2520Voss%2CElara%2520Voss&amp;date=today%205-y&amp;geo=Worldwide&amp;gprop=web">Google Trends</a>.</figcaption></figure><p>The problem is not that the technology lacks power. The problem is that without constraints, power defaults to the average. And the average, in creative writing, is the one thing nobody sets out to produce.</p><h3>My Attempt to Solve It</h3><p>There are tools that address parts of this. <a href="https://sudowrite.com/">Sudowrite</a> offers a Story Bible for tracking characters and locations, and its Muse model is fine-tuned on fiction. <a href="https://www.novelcrafter.com/">NovelCrafter</a> maintains a structured database called the Codex, which auto-injects relevant character and setting details into each prompt. <a href="https://novelai.net/">NovelAI </a>tracks characters and settings through a “Lorebook” system across long-form stories. <a href="https://www.squibler.io/">Squibler</a> adds a structured drafting workspace with elements and chapter-oriented generation tools. These tools solve real problems, and they add more structure than simple prompt-in text-out workflows. But they still center on generation-time context and drafting acceleration. The gap I kept hitting was workflow governance: explicit quality gates, constitution-style compliance checks, and cross-chapter course correction as first-class steps.</p><p><strong>Author Kit is what I built to work at that other layer.</strong></p><p>It is an open-source toolkit for structured AI-assisted manuscript writing. It uses commands (or prompts, depending which AI model you use) organized into a sequential workflow. You define your book’s concept before you outline. You outline before you draft. You review each chapter against your own standards before moving on. If something drifts, you catch it. If something needs to change, you see the impact before you commit.</p><p>The philosophy will sound familiar to software engineers. In the past two years, spec-driven development has emerged as the antidote to what people call vibe coding: prompting an AI, hoping for the best, and manually fixing whatever comes out. Author Kit applies the same principle to writing. SDD is to vibe coding what Author Kit is to vibe writing. You define the spec, generate from it, and verify the output. The spec is the source of truth. The prose is the generated artifact.</p><p>Two artifacts sit at the foundation of every Author Kit project.</p><p>The first is the <strong>constitution</strong>. It is a style guide with enforcement. Before writing a single word of prose, you define the rules: point of view, tense, emotional register, prose standards, what counts as too much description, what kind of humor is allowed, what is off limits. Here is an excerpt from a real one:</p><pre>### Prose Standards<br><br>- DO: &quot;The door creaked open.&quot;<br>- DON&#39;T: &quot;The ancient, weathered door slowly and painfully creaked open.&quot;<br><br>No hedging language. Say what is true, flag what is uncertain,<br>skip the middle ground of &quot;it could be argued that perhaps.&quot;</pre><p>Every chapter review checks compliance against these rules. If the constitution says first person present tense, a drift into past tense in chapter seven gets flagged. If it says no adverb-heavy prose, “she walked slowly and carefully across the room” gets flagged. The constitution is not a suggestion. It is a contract between you and the generation process.</p><p>The second artifact is the <strong>concept file</strong>. Before outlining, before building your world, you define what this book is. Premise, themes, audience, scope, success criteria. Then you go through clarification rounds: the AI asks targeted questions about under-specified areas. Not <em>“what does the character look like?”</em> but <em>“the concept implies a nonlinear timeline with a first-person narrator — is the narrator aware of the nonlinearity, or is the reader discovering it independently?”</em> Each answer gets encoded back into the concept. By the time you are done, the concept is specific enough to generate from.</p><pre>/authorkit.conceive A story about an ordinary sailor in the U.S. <br>war of independence against the British, but who fought with Britain.<br>As they get defeated, he tries his chance in America under another name,<br>risking being discovered.<br>Remember to do it in a factual, historical accurate manner.<br>Follow a style similar to Gustave Flaubert.</pre><p>This matters because of what we just discussed. Without these artifacts, the AI draws on the full distribution of its training data: every book it has ever seen, every style, every genre, every convention. The output is the statistical average of all of them. With a constitution and concept in place, the output space contracts. The AI is no longer generating “a story.” It is generating this story, for this audience, in this voice, with these constraints.</p><p>The constitution and the concept are the difference between <em>“write me a novel”</em> and <em>“write me a first-person literary thriller set in 1890s London, dry humor, no adverbs, chapters under three thousand words, centered on a female astronomer and the cost of obsession.”</em></p><p>Both prompts produce prose. Only one produces a differentiated book.</p><p>If I had written a constitution for the cave story, it would have included a rule: <em>“Only living organisms duplicate. Objects, vehicles, and infrastructure are not affected.”</em> That sentence, loaded into every chapter’s generation context and enforced during review, would have prevented the car problem entirely. Not because it would have made the AI smarter, but because it would have made the output space smaller, closer to what I actually wanted.</p><h3>What Comes Next</h3><p>This article defined the problem my approach to AI-assisted authoring is trying to solve: unstructured AI generation defaults to the statistical average, and no amount of prompt variation fixes it. It introduced the first layer of the response: defining intent through constitution and concept before generation begins.</p><p>But intent alone does not hold across twenty chapters. My cave story proves that: you need to be able to track the evolution of the story as the world you operate on grows. Many copies of the same person are impossible to track with a concept file alone.</p><p>This article is the first of a series of four that introduces Author Kit. The next one shows how intent becomes structure. World-building in Author Kit is not static. It is an entity database that tracks every character, location, and system across every chapter. The outline is not a suggestion. It is a structural contract. After that, a third article covers what happens when the structure itself drifts or and the author wants to pivot. I will also have a final article that reflects honestly on what works, what does not, and what remains to be solved.</p><h3>Trying Author Kit</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7n7nsDJhzEvAhXJZY9srrw.png" /><figcaption>Source: Image generated by the author</figcaption></figure><p><a href="https://github.com/mazemerize/author-kit">Author Kit</a> is open source. However, you will need an AI Coding agent at your disposal (Claude Code, Codex or GitHub Copilot) and a Python environment with uv. I acknowledge that for many writers, this can be quite tech oriented. Later, if any interest, a better UI/UX could be developed into a product.</p><p>Install Author Kit by running:</p><pre>uv tool install authorkit-cli --from git+https://github.com/mazemerize/author-kit.git</pre><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0b02323f618a" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>