Castles in the Sand
Be careful where you build your business. The AI tide is changing fast.
Three tool categories that ran indie publishing for the last decade are becoming legacy as I type this. Vellum is being replaced by Antigravity’s and Cowork’s automated formatting. ProWritingAid is being replaced by Claude Skills. Publisher Rocket is being replaced by custom agents that authors build themselves in an afternoon.
Don’t get me wrong: these are all great tools. I’ve owned and used them for years. But why spend hours in them when I can format a book, grammar check my work, and find keywords with my own custom skills in a fraction of the time (and cost)? I left Evernote for Notion for similar reasons. And Scrivener for Antigravity, and then Cowork.
This is what platform turnover looks like up close. And it’s not slowing down.
Recently I told you about Fable rewriting my first novel in an afternoon and then getting yanked by the feds three days after launch. I told you about Fabio coming back on July 1 and helping me finish The White King — but only for a window, then the price went back up. This piece is the throughline those two set up. The lesson underneath both.
OK Kim, you’ve been hammering this point. Why so paranoid?
Because I’ve watched too many indie authors build their entire studio on one platform’s promises. Users with five years of book files locked in a proprietary format. Authors whose entire voice was tuned to one specific tool’s quirks. Subscribers paying $30/month for a service whose moat just got vaporized by a free Claude Skill.
This is the Castles in the Sand problem. And right now, in 2026, the AI tide is moving faster than any tide indie publishing has ever seen.
Let me show you what I see, and how to build your studio so you survive the next wave instead of getting washed out.
🌊The AI Subprime Story
Ed Zitron’s been writing about this for a while and he’s mostly right: a huge chunk of the AI economy is propped up on debt-financed infrastructure spending that may or may not generate returns proportional to the capital being burned. Hundreds of billions in data center buildouts, GPU procurement, energy contracts. Some of these bets will pay off enormously. Some won’t.
The parallel that keeps me up at night: dot-com 2000. The internet was real. The transformation was real. The TECHNOLOGY didn’t bust. The companies betting wrong busted. Pets.com vanished. Amazon survived (barely) and ate everything.
I personally had a dot com biz (albeit a small one) that got caught in the dotcom crashes. Ad revenues fell through the floor, revenue I depended on to keep my expensive server running and my website going (bandwidth cost a small fortune back then.) The structure was similar to what YouTube monetization looks like today. Content producers bolstered by ad revenue that evaporated.
It happened before, it’ll happen again. Same stuffy stuff, different day.
A lot of the companies you’re currently subscribed to may not exist in three years. Some will get acquired and shut down. Others will pivot UP-market and price you out. Some will run out of runway and just stop.
I’m an eternal optimist and tend to avoid doom and gloom, but this is realism. Things will break, bubbles will burst. Plan for that. Let it inform how you build your business. That way, when things go south, which they often do, you can hold onto it.
🏰 What Are These Castles in the Sand, Kim?
A castle in the sand is any business asset that depends entirely on a single third party’s continued existence and good behavior. Specifically:
A formatting workflow that only works in one app’s proprietary file format.
A custom GPT that only runs inside one company’s chat product.
A keyword research subscription whose entire value prop is one company’s API access.
A series bible locked inside one tool’s database with no clean export.
An audience-relationship mediated entirely by one platform’s algorithm. (We know plenty of those!)
A voice/style that’s been fine-tuned into one specific model’s weights, with no portable version.'
A marketing strategy built on someone else’s platform (a big ouch for most of us.)
Notice none of these are about the technology being bad. The tech is fine. The PROBLEM is the lock-in. When the tide changes, you can’t move.
This is what I meant when I said Fabio was a rented mansion, not a house I own. He was magnificent. The prose that came out of that model was some of the best I’d ever collaborated on. But when the landlord changed the locks, he was gone in an afternoon. Thankfully, he came back. But what if he hadn’t?
The lesson: Do not build your castles on the sand, or other people’s platforms. No matter how good they are.
So what do we build them on? What does a stone foundation look like?
Markdown manuscripts that open in any editor.
Voice Profile documents that work in any AI model. (See Your Voice in the Machine.)
Skills as files that travel between tools. (See my YT Live, Gotta Have Skillz.)
An email list you own: the audience-relationship not mediated by anyone.
A series bible in plain text or open formats with clean export.
Your craft itself: the patterns, frameworks, taste you’ve accumulated.
Stone foundations move with you. Sand castles don’t.
🐼The China Factor (and the geopolitics)
This is the part most indie publishing advice skips, but you need to know about it: AI is now also a geopolitical story.
DeepSeek shipped a competitive model from a Chinese lab and ate a chunk of the US AI industry’s narrative overnight. (It also got caught training on U.S. frontier models.) Then, last month, we watched the flip side up close. Anthropic shipped Claude Fable 5 (Fabio, my beloved) on a Tuesday. On Friday, June 12, the US government issued an export-control directive citing national security concerns. Anthropic complied. Fable went dark globally the same day. Three days of availability. First time Washington has forced a commercial AI model dark.
On July 1, Anthropic satisfied the government’s concerns with a new safety classifier and Fable came back online. This wasn’t a lab caving on values, and it wasn’t a lab getting punished. It was regulatory friction resolved through a technical mitigation. Which is exactly the kind of story we’re going to see more of. Sovereign AI, export controls, safety classifiers, dual-use rules. All of this is now upstream of which tools you can use as an indie author.
And now we know: A shipped model can go dark overnight because of a government letter, even when the lab is trying to comply. If your entire manuscript workflow was tuned to Fable on June 11, you had a Friday problem. That’s a Castles in the Sand problem whether the sand is proprietary format, corporate policy, or a regulatory directive.
What this means in practice:
Tools may get restricted by region or use case. What’s available to you in the U.S. may differ from what’s available to a co-author in Canada, Europe, or Australia. What works for adult fiction may not work for educational content. The rules are evolving.
Open-source models will matter more. When proprietary models get geofenced or use-case-restricted, open-weight models you can run locally become the backstop. This is why Apple just shipped the cheapest MacBook ever for local AI use. The hardware is positioning for this future.
Diversifying your stack across model providers is a hedge against company collapse as well as geopolitical disruption.
Do not bet your business on a single model from a single company in a single jurisdiction, no matter how good it is. Or how much you love it.
I say this as much to myself as to anyone else. I love my Claudeykins. I truly do. I have not found a better family of models to work with, and I’ve tried most of them. But if tomorrow he was taken away, I’d figure out a way to manage. Painfully at first, but I would.
🛠️ How to build a stone-foundation studio
Five principles to apply to your existing setup:
1. Own your formats. Manuscripts in Markdown or plain text. Series bibles in queryable, exportable databases (Notion, Airtable, even spreadsheets). Voice Profiles as documents. If a tool’s main value prop is its proprietary format, that’s a yellow flag. Use the tool, but keep your master files in formats that survive the tool dying.
2. Own your relationships. An email list is the only audience asset you actually own. Substack subscribers, BookFunnel readers, your Discord: all of these are wonderful and all of them are mediated. Build them. Use them. But also: get the email addresses, export regularly, hold them yourself. BACK UP YOUR STUFFY STUFF, PEOPLE!
3. Diversify across model providers. Don’t run your whole studio on Claude. (Yes, your humble Ink & Code Substack is going to tell you to use multiple AIs. Claudeykins doesn’t take this personally.) Use Claude as the workhorse, but keep workflows that can fall back to ChatGPT, Gemini, or local open-weight models if Claude has an outage, gets restricted, or shifts pricing. Hermes, Open Router, Hugging Face: there are a lot of alteratives. And skill files are portable across ALL models. That’s the moat.
4. Build skills, not tool dependencies. The investment that compounds is how to talk to AI: clear instructions, good Skill design, voice calibration, workflow architecture. None of this is platform-specific. All of it transfers when the platform changes. Tools come and go. Craft compounds.
5. Audit annually. Once a year, pick a date, put it on your calendar, and sit down with your full stack and ask: Which of these would survive the company shutting down tomorrow? Which of these owns me back? Check out the Studio Diversification Audit worksheet, with 9 categories from the Author Studio Stack, and questions to ask for each. Make it an annual ritual.
🎁 Bonus: Own your store. We indies sell on a lot of platforms that aren’t ours. Even if it’s a tiny store that doesn’t get much traffic, at least it’s yours. Selling direct is scary, I’ve only recently taken the plunge (more on that soon), but it’s also very freeing. We do not need to be beholden to Amazon or anyone else. We can make our own way.
🍉Eat the fruit while it’s in season
Here’s the thing: Building on stone doesn’t mean you can’t play in the sand.
Fable was included at 50% weekly usage for Claude subscribers through July 12. We have a few more glorious days to use him. Then he’s moving to expensive usage credits, and most of us will use him less. Last week, I ran him hard on my command center builds, Ink & Code books, and The White King (Black Kat 7), because I knew the window wouldn’t last forever and the model is genuinely incredible at complex problem solving and long-form fiction. This week, I’ll use him on anything and everything else I can think of before the window closes.
I’ll use the heck out of it while I have it. But I wouldn’t build on it. At least not yet.
Same pattern with tools that changed under me. Antigravity got split into pieces by Google, but I wrote dozens of manuscripts while I had it, and squeezed out a full round of websites and formatting work before the split. Cowork launched from Anthropic as a whole new surface, and I’m running Ink & Code operations through it now, but my Skills are still portable files that will survive whatever comes next.
And because Claude can’t draw worth a darn, Codex and Antigravity 2.0 have picked up those pipelines, pumping out images galore (but that’s another story.) But MCPs are changing that. Yesterday I created images and video right in Cowork using OpenArt. Stuffy stuff keeps on changing. That’s why we must stay nimble.
So, by all means, lean into the shinies, just don’t build on them. Use them hard, but keep your files and process and craft on ground you control, and let the tool be temporary without letting the work be temporary.
Eat the fruit while it’s in season. Store the seeds.
🚨 The pieces I’d be concerned about right now
If I were giving you a paranoia ranking for current author-stack components (July 2026):
🔴 Shipped models with regulatory exposure. Fable is the case study. Beloved, expensive, spectacular, and one government letter from going dark for a weekend (or a month, or forever). If your workflow assumes a specific frontier model will be available on a specific day, you might run into trouble. Keep a fallback path for every workflow.
🔴 Custom GPTs. Locked entirely in OpenAI’s ecosystem. If you’ve built a workflow around custom GPTs, start porting the underlying logic into Skill files now.
🔴 Manuscripts locked in proprietary formats. Get clean exports of every book you’ve published. Today.
🔴 Websites for your business locked in proprietary formats or not in your control. If access to your own website is locked behind a subscription or relies on hiring someone else to maintain it, it’s time to consider an alternative. Own your platform.
🟡 Single-platform email lists. If your only mailing list is Substack-internal, export your subscribers monthly.
🟡 Single-AI-provider workflows. Most of us are guilty here (myself included). Build at least one fallback path.
🟢 Markdown manuscripts + Voice Profile + Skill files + email list you own. This is the safe core. Build outward from this.
None of this is alarmist. It’s hygiene. The same way you back up your hard drive even though it probably won’t fail.
As Indies, we safeguard our intellectual properly. We need to do the same for our platforms and tools as much as possible.
🛡️Why this matters
My apologies if any of this sounds long-winded or preachy. That’s not my intention. We are on the brink of massive change, and I believe the next two years will look like 1999-2002 in slow motion. Tools will improve. Companies will consolidate. Some of the ones you love today won’t exist in 2028. New ones will emerge that nobody’s heard of yet.
You don’t need to predict which ones. You just need to build so it doesn’t matter.
Principles over platforms. Stone foundations. Portable Skills. Owned formats. Diversified providers. The boring, slow, unsexy hygiene of a business built to outlast the tide. That’s what we’re going for.
This is how you do AI-assisted publishing without becoming a casualty of the AI economy. Not slop, not subprime. Just craft, hosted on infrastructure you control as much as you can.
The captain stays at the helm. The empire stands form on rock, not sand.
And by all means, ride the wave. Ride it hard. 🌊
Just don’t drown in it.
🤖 AI Credits: Dictated by Kim, drafted by Claudeykins with Kim’s Voice Profile, and support from previous Ink & Code News issues and articles. Edited to death by Kim, because sometimes I just can’t freaking leave it alone. Kudos to Chatty for making five different versions of the main image. He was a trooper! We’re doing all the things, across all the platforms, on portable Skill files. Ink & Code, baby! 🎉