All Things AI · Season 3 · What You Can Build · Episode 15
Not a paragraph you build around. The whole piece, structured, produced from your own material, ready for you to correct rather than to begin. The difference between those two outputs is the difference between a tool that helps and a tool that changes your week.
A paragraph is not a draft
The common experience is asking for something and getting three competent paragraphs that could open almost anything. You then write the actual piece, using them as a warm-up.
That is not nothing. It is also not what was promised, and it explains why so many agents conclude the whole category is oversold. They were sold a writer and given a starter.
A full draft is different in kind rather than in length. It has a structure, it makes specific points in a deliberate order, it uses your material rather than generalities, and it ends properly. What you do to it is correction rather than construction.
Why the short output happens
Almost always because the request was short.
Write me something about this contains no structure, no audience, no purpose, and no material, so the model supplies all four from the average of everything it has read. The average is generic by definition.
The models are capable of far more than most people get from them, and the gap is nearly always the specification rather than the model. This is the single highest-leverage thing to fix and it costs nothing.
What a request that produces a real draft contains
Five things. Leave any of them out and the model fills the gap with an average.
- Who it is for, specifically. Not homeowners, but someone in a particular situation with a particular concern.
- What it is supposed to do. Inform, persuade, reassure, or prompt a decision. These produce genuinely different pieces.
- The structure, in order. What comes first, what comes next, and what it ends on. If you do not specify this you are asking the model to guess the shape of your argument.
- The material. Your notes, your file, your actual specifics. Without this it writes from general knowledge, and general knowledge is what makes writing sound like nobody.
- The length and the register, stated plainly.
Where the structure should come from
You, and this is the part worth protecting.
Handing over the structure is handing over the argument, and the argument is the thinking. A model asked to decide what order your points go in will produce a reasonable order that is not yours, and the piece will read as competent and impersonal in a way that is hard to diagnose afterwards.
Spend five minutes on the outline. Five bullets is enough. That five minutes is what makes the resulting draft yours rather than a well-written piece about your topic, and it is also the difference between correcting a draft and rewriting one.
Running it as a full pass
Ask for the whole thing at once rather than section by section.
Building a piece in fragments produces a piece that reads like fragments. Each section is written without knowledge of the others, transitions do not exist, and the same point gets made three times because nothing knew it had already been made.
One pass produces something coherent. It will be less polished in places than a carefully assembled version, and it will hold together, and holding together is much harder to fix afterwards than a rough sentence is.
Then change one thing at a time
The correction habit matters as much as the drafting.
When a beat is wrong, send back the beat rather than the piece. Rewriting the whole thing to fix one paragraph reliably introduces two new problems elsewhere, and it is how a draft that was close ends up worse after three rounds of improvement.
This is the same discipline that makes a model trained on your own writing useful, and it applies here even without one. Small targeted corrections converge. Wholesale rewrites oscillate.
Which model to use for this
Long-form drafting is one of the areas where open models running on your own machine are genuinely competitive, which makes this a good candidate for running free and locally.
GLM in particular is strong at sustained structured writing, which is exactly the shape of this job. Running it locally also means you can put your actual material in without thinking about it, and your actual material is what separates a real draft from a general one.
Where a paid service still helps is on the hardest reasoning, and most drafting is not that. Try it locally first and you will likely stop paying for this particular task.
Where this goes wrong
Four failures, and the first two produce almost all the disappointment.
Not supplying material. A model with no specifics writes generalities, and generalities are what make content forgettable. If you did not give it anything only you know, it cannot produce anything only you could have written.
Letting it choose the structure. Covered above, and it is the subtle one, because the result is competent enough that the problem is hard to name. The piece is fine and it is not yours.
Building in fragments. Produces repetition and no transitions, and both are harder to fix than they would have been to avoid.
Accepting the first draft because it is good enough. Good enough published repeatedly is how a voice erodes, and the correction pass is short.
Keeping a shape you reuse
Once a piece comes out well, the valuable thing is not the piece. It is the request that produced it.
Save it. The audience line, the purpose, the structure, the register, the length. Next time you need something of the same type, you are changing the material rather than starting from nothing, and the second piece takes a fraction of the time the first one did.
After a few months you have a small set of these for the things you actually produce: the explainer, the update, the answer to a common question, the piece that introduces something. Each one is a shape that reliably produces work you are happy with, and building that set is most of what separates someone who gets a lot out of this from someone who gets a paragraph.
This is also the cheapest possible version of building a tool. It is a text file of requests that work, and it delivers a large share of the benefit before you automate anything at all.
What it grows
Throughput on everything that needs words, which in this business is very nearly everything.
The constraint on how much an agent publishes has never been ideas. It is that turning an idea into a finished piece takes an hour you do not have, so most ideas stay unwritten and the ones that get written are the ones that happened to coincide with a free evening.
Removing the hour changes what gets made. Not by making you a faster writer, but by removing the reason most of what you thought of never existed.
What it costs
Free, if you run it on an open model on your own machine, which is where this particular job belongs. On a paid service it is small per piece, and the volume this enables is exactly what makes running it locally worth setting up.
Questions agents actually type
Why does AI only give me a few paragraphs instead of a full draft?
Almost always because the request was short. Without an audience, a purpose, a structure, your material, and a length, the model supplies all five from the average of everything it has read, and the average is generic by definition.
Should I let AI decide the structure of my writing?
No. The structure is the argument and the argument is the thinking. A model will produce a reasonable order that is not yours, and the result reads as competent and impersonal in a way that is hard to diagnose later. Five bullets from you fixes it.
Is it better to write section by section or all at once?
All at once. Sections written separately have no transitions and repeat each other, because nothing knew what the other sections said. A single pass is rougher in places and coherent overall, and coherence is much harder to add afterwards.
How should I edit what AI drafts?
Send back the beat that is wrong, not the whole piece. Wholesale rewrites to fix one paragraph reliably introduce new problems elsewhere, which is how a draft that was nearly right gets worse over three rounds of improvement.
Season 1 is Claude Code for Real Estate Agents, filmed and on its own hub. All three seasons: AI for real estate agents.