Write the Words That Matter, Delegate the Rest

Here are two facts that sit awkwardly next to each other. In 2026, large language models can generate not just working code but the documentation that explains it. And yet many developers, given that ability, explicitly tell the model not to. One writeup making the rounds, titled “2x, not 10x: coding with LLMs in 2026,” puts it bluntly in its instructions to the AI: “Never write READMEs, docstrings, or comments. I will write those myself later. And yes, I really mean this.”

That last line does a lot of work. It reads like someone who has been ignored before. And if you’re a non-technical person trying to figure out how to write with an LLM, that single sentence is more useful than most advice you’ll find.

What the coders figured out first

Software developers have been the guinea pigs for AI-assisted writing, whether they wanted to be or not. They’ve had these tools in their daily workflow longer and more intensely than almost anyone else. So when they start drawing lines around what they’ll hand off and what they’ll keep, it’s worth paying attention.

The line they keep drawing is interesting. They’re happy to let the model produce the machinery. They’re much less happy to let it produce the explanation of the machinery. READMEs, comments, docstrings — these are the parts a human reads to understand intent. Why did we build it this way? What should you watch out for? What was the tradeoff?

An LLM can produce something that looks like an answer to those questions. It reads the code and describes what the code does. But describing what something does is not the same as explaining why it exists. The why lives in the author’s head, and no amount of pattern matching pulls it out.

The writing parallel

Swap “code” for “content” and the same split applies to your work. The parts of your writing that are structural, repetitive, or derivable from something you’ve already decided? Good candidates for delegation. The parts that carry your judgment, your reasoning, your particular take? Those are the parts you write yourself, and you should be as stubborn about it as that developer was.

In practice this might mean the model drafts your product descriptions from a spec you wrote, but you write the paragraph explaining why you changed the pricing. It might mean the model reformats your notes into a clean outline, but the argument inside it is yours.

Plan first, generate second

The other lesson from the developer world is about sequencing. Addy Osmani, describing his own workflow going into 2026, names the common mistake directly: jumping straight to generation with a vague prompt. His fix is to brainstorm a detailed specification with the AI first, then outline, then generate.

Non-technical writers make the exact same mistake, just with different words. “Write me a blog post about customer retention” is the vague prompt. What comes back is technically a blog post and emotionally a shrug.

The better move is to spend your first few exchanges building the plan rather than the product. Who is this for? What’s the one idea they should walk away with? What are three examples I already have in mind? What should this absolutely not sound like? By the time you ask for a draft, the model isn’t guessing. It’s assembling something you’ve already designed.

This is also why example-driven prompting works so well. Showing two paragraphs you like teaches the model more about your voice than a paragraph of adjectives describing your voice. Examples are specific. Descriptions are not.

Flip the direction

One pattern I like comes from a discussion of LLM architecture in 2026, and it inverts the usual setup. Instead of asking the model to produce text, you hand it text you already wrote and ask it to check the text. Here’s my article. Here’s my list. Make sure everything on the list is covered.

The model becomes a reviewer rather than a ghostwriter. You keep authorship, and you get a second pass that never gets bored on paragraph fourteen. For anyone nervous about AI writing replacing their voice, this is the gentlest possible entry point.

The models are not the variable

You’ll see a lot of comparison content about which model to pick. GPT-6 and MiniMax M3 are among the leaders, and modern systems can hold tens of thousands of coherent words together — a far cry from early models like GPT-1, which would start producing nonsense after a handful of sentences.

That capability is real, and it’s also not your bottleneck. The gap between a mediocre result and a good one almost always comes down to how much thinking you did before you typed. The developers who landed on “2x, not 10x” weren’t complaining about their tools. They were being honest about where the gains actually come from.

So write the plan. Show examples. Keep the sentences that only you could write. Let the model do the rest, and mean it when you tell it where the boundary is.

🕒 Published: September 18, 2026