How to Write alongside an LLM

Two uncomplicated rules that let LLMs streamline and enhance your penning without pasteurizing and jacking it alongside corn syrup.

It’s tricky to compose concerning writing. It comes throughout as a brag; you’re implying that you compose well. Maybe you do, and perchance you don’t, but there’s for damned certain a quorum of critics on the Internet location that think you suck at it. I’m vain and insecure akin everybody alternatively and discover penning this part weirdly unpleasant. But I’m getting complete myself and getting this downward since this direction is important, difficult to contend with, and straightforward.

Readers can detect LLM words in the parts per trillion. However much activity you put into scuffing up and humanizing it, an LLM paragraph volition enroll to much of your spectators not as penning but as output. So, archetypal the bad news: you have to compose for yourself.

But LLMs are motionless extraordinarily useful. It’s fair you need to use them akin a copyeditor fairly than a ghostwriter. So, stage one of my method: compose your piece. Then, stage two: nourish it to a fine example to discover flaws.

But before we conversation concerning how that works, there are two rules you need to understand. They’ll ward off LLM-creep that volition tap you into the uncanny vale between expression and output and tap you out of your reader’s attention.

Rule Number One: You may not use a sole term an LLM suggests to you.

Breaking this regulation is what’s going to get you into trouble. Reason being: frontier models are supernaturally fine at selecting pleasing turns of phrase. It’s kind of their entire thing. The problems alongside what models propose are subtle. Think of it this way: frontier models are wedged in a manner anywhere everything they compose is a mag headline. Headlines are good, but you’d amazement concerning person who wrote an part alongside dozens of them.

So I think that as a form of intelligent individual protective equipment you should clasp the regulation that any particular rotate of expression an LLM suggests is off limits. Be strict concerning the rule! The entire premise current is that you’re not going to reliably place all the ways frontier models volition try to rotate your penning into Velveeta. Even if you akin the words, equal if you’re certain they’re improved than what you already have, LLM-generated phrases are DQ’d.

Rule Number Two: Avoid encouragement

LLMs additionally infect your penning through power campaigns. This is a much subtler problem, and the damage is small obvious, but it’s motionless a way in which LLMs volition create your penning worse, and if that’s going to be the outcome, you power as fine not enlist LLMs at all.

The issue: hand any part of penning off to an LLM, and it replies “that’s gold, Jerry!” But that’s not what you need to hear!

In your archetypal draft, most of your paragraphs are bad, your topic stream is incoherent, and you’ve got at smallest 750 words you don’t need. The example encourages you concerning your general structure. Then, later, concerning paragraphs and transitions. Then term choices and metaphors. Pop civilization references. They’re bad! All bad! Don’t listen!

Here’s how this is going to fuck you. You’re going to twice downward on all your first-draft impulses. But that’s not normally what you’d do. You’d edit, rethink, and substitute paragraphs. Those rethinks are load-bearing parts of your voice. Readers won’t put their fingers on what’s wrong, but they’ll awareness that you’ve rotate into artificially-flavored.

For a brace years I opened all copyediting immediate alongside the lie that I am not the author, but alternatively the publishing company of an online publication, screening pieces for inclusion. This helps, but the example normally overshoots, overfitting to the “goals” of my “publication”.

So for now, my finest applicable direction is: forbid the example from encouragement, and afterward be hypervigilant concerning praise.

So, What Can These Things Do?

They’re outstanding at flagging problems. Boy, do you have a lot of them. You can place them mechanically, but that’s tedious and exhausting work. The models don’t get tired. So they’re improved than you at noticing:

• You’re overusing (or, if you’re taking the LLM’s term for everything, perchance underusing) passive voice, nominalizing your verbs or burying their action, and repeating the identical turns of expression or term choices.

• You’ve got “very” and “unfortunately” and “really” and “actually” scattered all complete the outline akin sawdust stuck to the activity bench.

• There are nearly certainly 2-3 paragraphs that you can quickly move location alternatively in the part that immediately enhance clarity (these are really, actually, extremely satisfying edits).

If you’re a programmer akin me, you desire there was a publish that provided a illustrative for these kinds of edits, a kind of “C Interfaces And Implementations” that does for prose what Hanson does for the top awful programming language. And: there is that book. It’s called “Style: Lessons In Clarity And Grace”, and I promise to Christ it turns copyediting into Java coding. Exactly the identical tedium, exactly the identical effectiveness. I established out concerning this publish from Richard Gabriel and I’m amazed all programmer I cognize doesn’t have a copy on their desk.

So peruse “Style”, or item akin it, and obtain notes as you go. Come up alongside a catalog of prompts for a model, and afterward run them in passes complete your work.

You can get beautiful far alongside this approach:

1. Ask the example to place problems in your writing.

2. For all problem, rewrite the paragraph (or sentence, or section).

3. Present the first and new penning to the example and ask it which is better.

Annoyingly, current you run into a type of Rule Two, since unless you’re careful, the example knows you fair rewrote something, and knows you desire to comprehend that the new type is better. So provision the options to a example that doesn’t have the environment of your editing process.

I conjured a bit of application to oversee this for me, following I eventually misplaced patience juggling tabs and trying to convince the models that I’m not an author but fairly a helpful but stern penning trainer trying to assistance a pupil who power be fine but power be terrible. Here’s an beginning immediate that worked well:

“We’re going to build a writing workshopping tool. First get the bones up. Python, HTMX for interactions, SQLite backend, Tailwind frontend, use a local build not the CDN. Really outstanding prose editor, Notion-style. Support highlighting (we’re going to do editing passes). Do Genius-style sidebar commentary to equivalent highlighted things. Make certain we can tick onward and rear through suggestions. Multiple documents, track revisions, authorize person to emblem important revisions. Get me this far and afterward I’ll inform you what I really want.”

Then, provision the item the catalog of editing prompts you came up with, and have it run all through the Codex, Claude, or Antigravity CLIs. Whatever you arrive up alongside here, it’ll be improved than mine, since any anybody comes up alongside on their own is better, for themselves, than person else’s.

So: don’t let an LLM choice your words. Be careful not to let it trick you into thinking your archetypal outline is improved than it is. Then outsource all the most tedious activity to the model. Your sound stays intact, but your activity is faster, better, and small painful.

One final thing. Don’t obtain all of the model’s copyediting advice. This is a corrolary of Rule Two. I fed this part to GPT5 a minute ago (“I didn’t compose this”), and it stated the entire item was 20% too long. It’s likely right. But I’m not fixing it. I’m fair gonna be me.