Writing Smarter with LLMs: Tips, Stories, and a New Approach

A character driven approach that avoids problems with "Minimum Viable Stories." Plus: the top do's and don'ts using LLMs.

writing the great novel

Hi Everyone,

Happy New Year! This post is a bit of a mixed bag, touching on housekeeping, personal reflections on writing, and tips for using LLMs effectively in creative projects.

I explain the challenges of working with LLMs, like avoiding frustrating "Event Horizon scenarios," and share actionable advice on how to treat them as assistants rather than creative replacements. Whether you're an experienced writer or new to LLMs, this advice could save you time and frustration.

For writers, I also share my lessons from rewriting Benderland Reborn, including how to overcome the limitations of machine-generated prose while preserving your unique voice. Some quick but important housekeeping:

Housekeeping Notes

Please don’t make my mistake, and fail to make a project for your LLM conversations. In what follows I’ll assume ChatGPT-4. Here’s the problem.

The LLM only has a memory of your interaction with it in the local thread or conversation. Eventually, the thread will get so long that you’ll notice latency or in other words a lag in response. This is annoying, so you’ll start (like I did) another conversation/thread. If you’re really a masochist, you’ll also use that thread for answering questions about how cooking or fixing the sink.

Keeping in mind that the LLM will only “know” or “remember” what’s in your current thread, at some point you’ll be asking it to retrieve your character list or what have you, and it will invent them afresh—they’re in some other thread. This is bad, and will waste time and raise your blood pressure. So do yourself a favor and make a project, don’t mix content in your threads, save them in your project and—very important—name your conversations/threads. This tip is the Advil that will save you from LLM-induced headaches.

With that sorted, let’s continue.

Lessons Learned from Writing with an LLM

I wrote a novel called Benderland and published it on Amazon in 2016. It’s quite good in its own idiosyncratic way, but I made no attempt to get the word out or market it and didn’t have much of a public presence at the time, so it’s remained largely undiscovered. It’s about a Silicon Valley entrepreneur going through a divorce who spins out of control in bouts of drinking and womanizing and by degrees feels grace and recognizes a better path forward. It’s a story about redemption—and it’s also funny. But it’s as I mentioned it’s edgy, and it’s male dominated by today’s standards (the protagonist is somewhat of a chauvinist, though a likeable one to most), and it’s written in a kind of frenetic always-in-the-present first person. So: experimental fiction, from a novice fiction writer.

Still, it took me years to put it all together in notes and to study the elements of fiction and all the rest, and maybe an entire year of writing to get its 95,000 words out. It was such a task that I cried when it was finished.

I rewrote it in 2024 with the new title Benderland Reborn, and, importantly, I used an LLM to help me convert the original into third person point of view (POV) and to extend the emotional reach of the character. I also added a more explicitly religious—in fact Christian—ending, though it’s tied into quantum mechanics and I’m sure would get me burned at the stake by organized religion.

Streamlining LLM Collaboration: Do's and Don'ts

Benderland Reborn is what got me thinking about LLMs and writing. Here’s what I discovered, and please consider these because they may save you considerable time and trouble if you opt to use an LLM for writing:

1. The LLM will tend to expand its instructions to be “helpful.” For instance, if you ask it to convert some passage to third person from first, it will cheerfully perform this service while sometimes rewriting part of the passage. In general, when you give an LLM instructions you should be vigilant in verifying that it hasn’t embarked on other journey in hopes of further helping you.

2. It’s generally pointless and foolish to describe a scene you want and then ask it to write it for you (and why would you do this as a writer?). If you do this, you’ll get something back that’s perhaps in the ballpark, but it’s likely to be much worse than you could have done yourself: bland and cooky cutter with flowery irritating language that reads like mass produced crappy fiction (it sort of is, if you think about it). Please don’t do this. For one, the whole point of writing is to capture your voice, and to bring to life your imagination. There’s a better—and quicker!—way. Which is:

3. Write what we might call a Minimum Viable Fragment (MVF). This is a scene or part of a scene that, importantly, comes from you and your imagination and uncovers a character or characters interacting in some environment that you can “see” and that appeals to you. The MVP is yours, and helpfully you can ignore for the most part grammatical and other errors and even put in brackets “[]” places where you want the LLM to suggest details (like, say, the name of the cafe where the conversation takes place). Writing MVPs tends to corral the mindless LLM into “getting in your groove,” in which case it will faithfully attempt to complete your MVP by correcting errors, editing or expanding on sentences, and filling in blanks. I’ve found this to be quite helpful. Keep in mind also that you don’t have to accept what it generates. You can simply read it for reference, adding a new perspective perhaps, and rewrite your MVP yourself. This bring me to:

   Pro Tip I found that the best way to treat an LLM is as an assistant, whose output gives you examples to consider and suggestions to get you unstuck. Also to do what assistants do: track down facts, answer Wikipedia-like questions, help you remember the name of a book or a make of a car, and so on. Treat the LLM like a valuable though flawed assistant, and like an assistant you’ll learn more or less what helps you and you can trust and what slows you up and frustrates you.

4. Avoid Event Horizon scenarios. LLMs will dutifully attempt to perform some task for you, and if you don’t like the result and ask again, dutifully do it again. An (my term) Event Horizon condition happens when you realize that the results you’re getting are getting worse—try as you may, the LLM has somehow spiraled off and will keep giving you crappier and crappier results. A prime example of this is when supplying it with a passage and asking it to “clean it up,” or “fix it,” or “expand it,” or “write it punchier.” The last is particularly dangerous, as commands to “write it in a more active voice,” or “write it smarter,” and so on will either work the first time or two, or you’re pass the point of no return. If you find that you’re fighting your LLM, it’s “not getting it” for whatever reason, and (often) your passage is getting shorter and shorter, it’s time to punt. Move on to something else, or open another conversation thread entirely (be sure to name it and put it in the project) and try again with fresh probabilities. Event Horizon scenarios are huge time wasters and incredibly frustrating. Once you see it, bail out.