How To Write With An LLM — A Closing Ward

Two easy guidelines that permit LLMs streamline and enhance your writing with out pasteurizing and jacking it with corn syrup.

It’s tough to jot down about writing. It comes throughout as a brag; you’re implying that you simply write properly. Maybe you do, and perhaps you don’t, however there’s for damned positive a quorum of critics on the Internet someplace that assume you suck at it. I’m useless and insecure like all people else and discover penning this piece weirdly disagreeable. But I’m getting over myself and getting this down as a result of this recommendation is necessary, onerous to argue with, and easy.

Readers can detect LLM phrases within the components per trillion. However a lot work you place into scuffing up and humanizing it, an LLM paragraph will register to a lot of your viewers not as writing however as output. So, first the dangerous information: you must write for your self.

But LLMs are nonetheless terribly helpful. It’s simply you want to use them like a copyeditor reasonably than a ghostwriter. So, step one in every of my technique: write your piece. Then, step two: feed it to a great mannequin to search out flaws.

But earlier than we speak about how that works, there are two guidelines you want to perceive. They’ll push back LLM-creep that may knock you into the uncanny valley between expression and output and knock you out of your reader’s consideration.

Rule Number One: You might not use a single phrase an LLM suggests to you.

Breaking this rule is what’s going to get you into bother. Reason being: frontier fashions are supernaturally good at deciding on pleasing turns of phrase. It’s form of their entire factor. The issues with what fashions counsel are refined. Think of it this fashion: frontier fashions are wedged in a mode the place every part they write is {a magazine} headline. Headlines are good, however you’d marvel about somebody who wrote an article with dozens of them.

So I believe that as a type of mental private protecting tools it is best to undertake the rule that any particular flip of phrase an LLM suggests is off limits. Be strict in regards to the rule! The entire premise right here is that you simply’re not going to reliably spot all of the methods frontier fashions will attempt to flip your writing into Velveeta. Even for those who just like the phrases, even for those who’re positive they’re higher than what you have already got, LLM-generated phrases are DQ’d.

Rule Number Two: Avoid encouragement

LLMs additionally infect your writing by affect campaigns. This is a a lot subtler drawback, and the harm is much less apparent, however it’s nonetheless a manner during which LLMs will make your writing worse, and if that’s going to be the end result, you would possibly as properly not enlist LLMs in any respect.

The challenge: hand any piece of writing off to an LLM, and it replies “that’s gold, Jerry!” But that’s not what you want to hear!

In your first draft, most of your paragraphs are dangerous, your matter move is incoherent, and also you’ve bought no less than 750 phrases you don’t want. The mannequin encourages you about your total construction. Then, later, about paragraphs and transitions. Then phrase decisions and metaphors. Pop tradition references. They’re dangerous! All dangerous! Don’t hear!

Here’s how that is going to fuck you. You’re going to double down on all of your first-draft impulses. But that’s not usually what you’d do. You’d edit, rethink, and change paragraphs. Those rethinks are load-bearing components of your voice. Readers gained’t put their fingers on what’s flawed, however they’ll sense that you simply’ve grow to be artificially-flavored.

For a pair years I opened each copyediting immediate with the lie that I’m not the writer, however as an alternative the editor of a web based publication, screening items for inclusion. This helps, however the mannequin normally overshoots, overfitting to the “targets” of my “publication”.

So for now, my finest sensible recommendation is: forbid the mannequin from encouragement, after which be hypervigilant about reward.

So, What Can These Things Do?

They’re glorious at flagging issues. Boy, do you’ve gotten plenty of them. You can spot them mechanically, however that’s tedious and exhausting work. The fashions don’t get drained. So they’re higher than you at noticing:

• You’re overusing (or, for those who’re taking the LLM’s phrase for every part, perhaps underusing) passive voice, nominalizing your verbs or burying their motion, and repeating the identical turns of phrase or phrase decisions.

• You’ve bought “very” and “sadly” and “actually” and “truly” sprinkled everywhere in the draft like sawdust caught to the work bench.

• There are nearly actually 2-3 paragraphs that you would be able to shortly transfer some place else within the piece that immediately enhance readability (these are actually, truly, very satisfying edits).

If you’re a programmer like me, you want there was a guide that offered a schematic for these sorts of edits, a form of “C Interfaces And Implementations” that does for prose what Hanson does for the best horrible programming language. And: there may be that guide. It’s known as “Style: Lessons In Clarity And Grace”, and I swear to Christ it turns copyediting into Java coding. Exactly the identical tedium, precisely the identical effectiveness. I came upon about this guide from Richard Gabriel and I’m shocked each programmer I do know doesn’t have a replica on their desk.

So learn “Style”, or one thing prefer it, and take notes as you go. Come up with a listing of prompts for a mannequin, after which run them in passes over your work.

You can get fairly far with this method:

1. Ask the mannequin to identify issues in your writing.

2. For every drawback, rewrite the paragraph (or sentence, or part).

3. Present the unique and new writing to the mannequin and ask it which is best.

Annoyingly, right here you run right into a variant of Rule Two, as a result of except you’re cautious, the mannequin is aware of you simply rewrote one thing, and is aware of you need to hear that the brand new model is best. So give the choices to a mannequin that doesn’t have the context of your modifying course of.

I conjured a little bit of software program to handle this for me, after I lastly misplaced persistence juggling tabs and attempting to steer the fashions that I’m not an writer however reasonably a useful however stern writing coach attempting to assist a scholar who may be good however may be horrible. Here’s a gap immediate that labored properly:

“We’re going to construct a writing workshopping software. First get the bones up. Python, HTMX for interactions, SQLite backend, Tailwind frontend, use an area construct not the CDN. Really glorious prose editor, Notion-style. Support highlighting (we’re going to do modifying passes). Do Genius-style sidebar commentary to match highlighted issues. Make positive we will tick ahead and again by recommendations. Multiple paperwork, monitor revisions, permit person to flag main revisions. Get me this far after which I’ll inform you what I really need.”

Then, give the factor the checklist of modifying prompts you got here up with, and have it run every by the Codex, Claude, or Antigravity CLIs. Whatever you provide you with right here, it’ll be higher than mine, as a result of no matter anyone comes up with on their very own is best, for themselves, than another person’s.

So: don’t let an LLM decide your phrases. Be cautious to not let it trick you into pondering your first draft is best than it’s. Then outsource all probably the most tedious work to the mannequin. Your voice stays intact, however your work is quicker, higher, and fewer painful.

One very last thing. Don’t take the entire mannequin’s copyediting recommendation. This is a corrolary of Rule Two. I fed this piece to GPT5 a minute
in the past (“I didn’t write this”), and it mentioned the entire thing was 20% too lengthy. It’s most likely proper. But I’m not fixing it. I’m simply gonna be
me.