Here's the situation I was trying to fix.
At some point in the last year I had four AI drafting setups running: a resume-and-cover-letter drafter, a meeting-follow-up drafter, a weekly-update drafter, and a long-form writing assistant for this blog. All four did roughly the same thing—take a rough input, produce a structured draft. All four were maintained. All four cost time.
Two of them survived. Two didn't. This piece is the honest comparison, because "should I use AI for drafting?" is the wrong question. The right one is "which drafts are worth delegating, and which are worth doing by hand?"
Why Drafting Is the Most Overrated AI Use Case
Three reasons, and they explain why so many drafting setups get built and abandoned.
Drafting Is Where the Judgment Lives
For a lot of writing, the draft is the thinking. Deciding what to say, what to leave out, and how to say it is most of the work—and it's the part that can't be delegated without losing the value. Asking a model to draft is often asking it to do the part you needed to do yourself.
Polished Output Hides Shallow Input
A model produces clean prose from almost any input. That's the seduction: the output looks like work. But if the input was thin, the draft is a well-formatted version of nothing, and the time saved on the draft gets spent fixing the substance—usually more time than writing from scratch.
The Templates Rot
A drafting system needs context: tone, format, recurring content. That context has to live somewhere, and it drifts. The setup that worked in March produces generic output by June, and you don't notice until you read a draft carefully and realize it doesn't sound like you at all.
The Two That Died
The Resume and Cover Letter Drafter
What it was: Feed it a job posting and my background, get a tailored draft.
Why it died: The drafts were technically correct and completely generic. They hit every requirement and had no voice. Worse, tailoring the draft to be actually mine took as long as writing it—because the tailoring was the work.
The deeper problem: I only apply to a handful of jobs a year. The volume was too low to justify any system. I was maintaining a drafter for a task I did four times a year.
What replaced it: A short, handwritten tailoring pass using the interrogation method from the resume rebuild piece. No system, no templates. Just the model as a question-asker, me as the writer.
The Long-Form Writing Assistant
What it was: Give it notes and an outline, get a draft of a blog post.
Why it died: This one took the longest to abandon because it was the most tempting. The drafts were readable. They had structure. They were also not mine—they had my outline and none of my thinking.
Here's the honest version: I'd spend forty minutes producing notes and an outline, get a draft, then spend ninety minutes rewriting it into something I'd publish. Writing from the outline myself took about eighty minutes. The "assistant" cost me fifty minutes and a slight loss of authorship over the piece.
The failure wasn't quality. It was that the drafting step was where the thinking happened, and I'd delegated it.
What replaced it: Nothing. I write the first draft myself, then use the model for specific, bounded tasks: tightening a paragraph, finding a better structure for a section, questioning whether an argument holds. Drafting happens in my head. Editing gets help.
The Two That Survived
The Meeting-Follow-Up Drafter
What it was: After a meeting, I dictate a messy summary; the model produces the coordination note—decisions, actions, open questions—in the standard format.
Why it survived: The task has three qualities that make drafting delegation work:
The structure is fixed. Decisions, actions, open questions. The model isn't deciding what the note should be—I am, by dictating. It's formatting.
The content is given. I'm not asking it to have ideas. I'm asking it to organize what I said.
The volume is high. Multiple meetings a week, every week. The time saved compounds.
The judgment—what counts as a decision, who owns an action—stays with me, because I'm the one dictating the raw material. The model handles the shape.
The Weekly-Update Drafter
What it was: I write a rough, unorganized list of what happened this week; the model turns it into a short status update in a consistent format.
Why it survived: Same profile as the meeting drafter. Fixed structure, content given, high volume. I'm not asking the model to know what I did. I'm asking it to arrange what I already know into a format I've defined.
The update used to take twenty minutes of staring at a blank field. Now it takes five minutes of typing and one minute of review. That's a real, repeatable saving.
The Pattern
Putting the four side by side, the split is clean.
What died:
Resume and cover letters
Long-form writing
What survived:
Meeting follow-up notes
Weekly status updates
The difference isn't the task type—all four are writing. It's three properties.
1. Is the structure fixed, or is defining it part of the work?
Resumes and blog posts require structural judgment. There's no single right shape, and choosing the shape is the thinking. Meeting notes and status updates have a defined structure. The model formats; you don't delegate the decision.
2. Is the content given, or is generating it the point?
Resume drafts and blog drafts require the model to generate substance from thin input. That's where it goes generic. Meeting and status drafts start from content I've already produced—I'm handing over arrangement, not ideas.
3. Is the volume high enough to justify a system?
Resumes: four a year. Blog posts: two to four a month. Meetings and updates: weekly, sometimes daily. The system only pays for itself at volume.
All three have to be true. Any one of them missing, and the drafting system costs more than it saves.
What I'd Do Differently
If I were setting up drafting systems from scratch:
Start with volume, not capability. The first question isn't "can AI do this?" It's "how often do I do this?" Below a certain frequency, no system is worth maintaining.
Test whether the draft is the thinking. For each writing task, ask: is the valuable part the words or the decisions behind them? If it's the decisions, don't delegate the draft.
Build the input, not the output. The surviving systems work because I produce good raw material and the model arranges it. The dead ones failed because I asked the model to produce the raw material too.
Delete faster. I kept the blog assistant alive for months past the point where the math was clear. The cost of a dead system isn't just its maintenance—it's the attention it takes to keep believing in it.
What Still Needs You
The parts of drafting that don't get delegated, in the systems that survived:
The raw material. I dictate the meeting summary and write the update notes. The judgment about what happened and what mattered is mine.
The structure definition. The format is a decision I made and maintain. The model fills it; it doesn't design it.
The review. Every surviving draft gets read and often edited. The model's output is a first pass, not a final one.
Knowing when the draft is the work. This is the core judgment, and it's the one that separates the setups that save time from the ones that just move it around.
The One-Task Version
If you're considering an AI drafting setup:
Pick one writing task you do at least weekly. Below that, the math probably doesn't work.
Write down the structure. If you can't define it in a few lines, the structure is the work—don't delegate the draft.
Produce the raw material yourself. Give the model content, not a topic. Arrangement, not ideation.
Run it for two weeks. Track the total time: producing input, reviewing output, and maintaining the setup.
Compare to doing it by hand. If the system doesn't clearly win, delete it. Keeping a dead drafter alive costs more than it looks.
Two setups survived, two didn't. The pattern wasn't about the tool or the task. It was about whether the draft was arrangement or thinking—and whether the volume justified the system at all.
The drafts worth delegating are the ones where the thinking already happened.
Better work first. More options next.
Make the workflow earn its place.
No feedback yet — submit the first.