A few months ago, I saw a Reddit post advertising some kind of AI development editor. The author claimed to have written novels, paid a development editor to review them, and been unsuccessful. Then their software-engineer husband vibecoded an AI tool in a weekend that supposedly produced all the same criticisms and revision suggestions, but for $20/month. Lots of skeptical Redditors objected that this was likely a scam. Nobody could actually find the tool.

But the premise stuck with me: here’s a hyper-specialized skill that maybe a hundred people in the world do really well, for a major premium, and a performant LLM might approximate it. So what does a development editor actually do for $3,000?

What You’re Paying For

I talked it over with Claude and with a couple of friends who have actual writing careers. What emerged is that “development editor” is a consulting job that requires good aesthetic judgment, connections, and self-promotion. For that sum, you can expect a well-read, finely-honed expert to read your manuscript closely (and potentially reread it), pore over key sections, perform a set of recognized structural analyses, and then write a long, thoughtful editorial letter. Afterwards, there’s usually a meeting where the editor walks you through the documents, hands you a marked-up copy, and brainstorms solutions.

Many authors believe there’s an implicit further transaction: that the $3k earns you a ticket out of the slush pile through the editor’s rolodex, if they like the work. The idea is that the development editor is vetting you during the process, and if they like what they see, they’ll connect you with serious agents and publishers. If that’s true, then it’s less a time-intensive editorial task than a gatekeeping function wrapped around a skill wrapped around a real time commitment.

But the reality is that most people doing this work don’t have an expansive network and can’t get you a contract. The Association of American Literary Agents actually enforces this separation: agents who offer editing must refund all fees if they later offer representation for the same manuscript. The skill just is discernment.

And it’s real work. For an 80,000-word novel, a developmental editor typically puts in 40 to 80 hours of active analytical labor: reading, rereading, building scene maps, drafting the editorial letter. At $3,000, that’s $40-75 an hour for highly specialized cognitive work.

And someone online claimed AI could replace it, right now, in early 2026, with a weekend’s work.

Put frankly: that sounds like bunk. But I thought it might be fun as a kind of test for how much a nerdy non-programmer could do. I missed the last bandwagon when everyone was building apps, but this, maybe, I could try. So I used AI to build Apodictic.

Why “Apodictic”?

I’d been writing about Arendt and judgment in my free time, and I kept trying to justify this experiment in terms of testing whether LLMs could exercise authentic aesthetic judgment. Early on, I came up with the working title “anotherpanacea’s development editor,” whose initials, APDE, sound a little like “apodictic.” Kant uses “apodictic judgments” to name necessary copulas: judgments like “Necessarily, bachelors are unmarried men,” where the “are” can’t be otherwise. That’s the kind of deductive logic we expect from computers. But somehow LLMs are giving us more than that.

The name also rhymes with something I discovered about how fiction works. A book can be wildly experimental, playing with form, genre, voice, and plot in ways that surprise, frustrate, and infuriate. That can be highly enjoyable, but only if the reader is primed for it. Otherwise, experimental fiction generates one-star reviews and never finds an audience. So even difficult fiction has to communicate its intentions to its readers somehow.

Development editors think about this in terms of a contract between the book and the reader: what are you telling the reader to expect? Do you deliver? Genres might actually offer something closer to apodictic inferences. Not every mystery is a Whodunnit (it might be a Howcatchem), but there’s always a reveal economy. By asking authors to articulate their goals, Apodictic doesn’t have to just read and wonder: it can test a novel’s putative thesis and genre self-identification against the conventions of that genre.

What the tool does, in practice, is take a first pass to infer a contract. The most performant large language models can do a surprisingly good job of this. It’s like asking the AI: “What am I trying to say?” When it gets something wrong, you correct it, and then (if it works) it tells you which parts of the manuscript are doing something you didn’t intend.

One design principle mattered more than any other: the firewall. Every AI writing tool I tried wanted to rewrite my prose. I didn’t need a co-writer. Apodictic diagnoses problems and identifies classes of solution. It never invents content: no new plot events, characters, dialogue, or imagery. You’re the writer. It’s the analyst. Without that boundary, the tool would just become another way for the LLM to take over your manuscript, and the whole exercise would be pointless.

What I Learned About Writing

Building the back-end reference files was the most fun part. It was also an opportunity to dust off all the fiction and narrative nonfiction guides I’d always wanted to figure out: the detailed craft skills a competent writer cultivates that I, as an academic and even as an editor, never had a professional reason to learn. The idea that every avid reader would make a good writer is like the idea that every avid magic fan would make a good magician.

Reverse-outlining, for instance, is probably obvious to anyone who does serious work on fiction. You can learn something like it in law school, and every logic professor has taught a version of argument diagramming. But thinking clearly about the reader experience and reveal economy, or measuring the proportionality of different narrative elements: that’s what you have to know to write a good novel, not just to read one. A beat map is a cool tool, and if you’ve reread a book a few times you could probably reconstruct one. But it’s not like such things come naturally. (At least not to me.) And once you know the moves of Save the Cat, you won’t be able to watch network television without seeing them everywhere, which is less a gift than a curse.

I’ve always been tempted by the idea that narratives make arguments much like philosophy papers do, and that’s pretty obviously reductive. But the most transformative thing I learned was simpler: development editors think in terms of a contract between the book and the reader. What are you promising? Do you deliver? Within genre constraints, fiction and narrative nonfiction flourish when the author has something to say and something to refute. These guardrails are part of what a development editor looks for and enforces. It’s a simplification, but for many authors and readers it’s a necessary one.

That idea reframed everything else I was learning. The difference between Happily Ever After and Happy For Now in romance isn’t a trivial genre tag; it’s a promise with real consequences if broken. Grimdark and hopepunk aren’t symmetrical moods; they make different contracts with the reader. I’d never thought about any of this before, and it’s already starting to change how I read.

What I Learned About LLMs

Here’s where it gets humbling. The first time I ran a simple “be a development editor” prompt on a recently published novel (Dungeon Crawler Carl), a five-line prompt got probably 60% of the insights that my much more elaborate tool produced. I was shocked.

Even more humbling: it turned out the simple prompt could do a great job without even reading the novel. I picked what I thought was an obscure time-travel novel from the late 80s, Leo Frankowski’s The Cross-Time Engineer, and Apodictic seemed to do well analyzing it. It was even insightful about the book’s misogyny. But so was the five-line prompt, because the misogyny is famous, and there are plenty of discussions of it on Goodreads and in reviews that ended up in the training data. Anthropic paid a $1.5 billion settlement for ingesting pirated books, too, so the model may have had direct access to the text. This is a known issue in benchmarks like the math olympiad, but I didn’t expect the model to waste weights on obscure mid-list science fiction from forty years ago.

Among many other things, what we’re seeing in LLMs is a spectacular project of knowledge compression. They’re big files, but they contain far more general knowledge than you’d expect.

Thankfully, performance drops precipitously with unpublished writing, which is, of course, what a development editor actually works on. A few other lessons:

The models are sycophantic. Everyone knows this on some level, but it’s really hard to get them to notice a criticism and sit with it rather than explaining it away. A lot of what I built was about making sure hostile perspectives survive the review process.

Structure matters, but less than you’d think. You can get 20-30% of the depth of analysis from Claude with a simple five-line prompt in incognito mode. The elaborate plugin structures the next 70%. The AI labs mostly think this kind of scaffolding is unnecessary: as datasets grow, they naturally incorporate writing expertise from the entire English-language corpus. I kept testing the simple version against mine and almost gave up when the simple version started getting really good.

Multi-model synthesis helps. You can get better results by asking the same question of Gemini, ChatGPT, and Claude, then asking one of them to synthesize all three answers. Gemini writes a little research paper on every question; ChatGPT applies clear structural thinking; Claude writes lyrically and thoughtfully, and has enough working memory to evaluate everything together.

Vibecoding Is Real

I am not a computer programmer. I can parse basic HTML and have some rudimentary database knowledge. I learned git for this project. But I was able to build an app version of the plugin and fully set it up. It’s not totally stable (the main instability is that it uses a janky backend cloud computer rather than more advanced hardware) but it works.

One thing about vibecoding: you can do it early in the morning and late at night. Mostly it’s testing, reviewing, asking for a specific fix, and then waiting around while the machine does the work. I built the basic thing on Google AI Studio in React in an evening.

A confession on models: Codex is faster and smarter at React programming than Claude, and it’s not close. Especially after the Pentagon fiasco, Anthropic has my loyalty, and Claude works best with my personal approach to text and writing. But Codex 5.3 is just more of a stickler for coding projects right now, at least for what I’ve been building.

What I Actually Learned

I started this project to test a dubious claim from the internet. What I didn’t expect was that building the tool would teach me more about writing than twenty years of avid reading had. AI makes crazy projects possible: a philosophy PhD with no programming background can prototype a working app. And the work that editors, writers, and critics do is going to change. Not because the models are better than human judgment, but because they’re good enough to sharpen it.

Try It Yourself

If you want to skip the tool entirely and try the bare-bones version, here’s the five-line prompt that gets you surprisingly far:

You are a developmental editor for fiction. Read the attached manuscript and write an editorial letter in Markdown. Identify what’s working structurally, what’s losing momentum or undermining its own impact, and provide a prioritized revision checklist. Include an adversarial stress test: inhabit hostile reader perspectives and identify what an uncharitable reader would attack. For each adversarial claim, commit to a severity rating before generating a counter-argument, and do not let the counter-argument reduce the severity. End with a “what not to touch” section.

And a three-line addendum that fixes a common failure mode where the model doesn’t actually read the whole manuscript and hallucinates the rest:

Read the complete manuscript before beginning analysis. If the text is long enough that your reading tool truncates or summarizes middle sections, read it in sequential chunks until you have covered every section. Do not estimate word count—count it or ask for it. Do not begin drafting the editorial letter until you can confirm you have read the final page.

Give it a try on something you’ve written. Start with the five-line prompt and see what it catches. If you want the structured version, Apodictic is here. It also runs as a Claude Code plugin and as a Custom GPT, both at no additional cost if you already have a Claude or ChatGPT subscription. And then tell me what it got wrong. The whole point of the adversarial layer is that these models want to be nice to you, and the tool is only as good as its ability to resist that impulse. If Apodictic pulled its punches on your manuscript, or praised something it should have flagged, I want to know. Drop a comment below or email me at anotherpanacea@gmail.com.