Never Let AI Write Your First Draft

There are two popular positions on AI writing, and they’re both wrong in ways that cost experts money.

Position one: let it write everything. Type “write a LinkedIn post about leadership,” post the output, repeat. The result reads like everyone else who typed the same prompt, because it is everyone else who typed the same prompt — the model’s median take on a topic, wearing your headshot. Readers have learned to smell it. So has Google, which has gotten aggressively better at burying it. Worse than either: it’s a confession. The one thing you have that no model has is the specific, scarred, paid-for judgment in your head, and publishing generic AI prose announces that you didn’t bother to use it.

Position two: refuse the whole thing on principle. Understandable — most of what position one produces deserves the contempt. But it means turning down the best junior staffer you will ever get for twenty dollars a month, out of a fear that only applies to using it wrong.

The resolution isn’t a compromise between the two. It’s noticing that both positions share the same wrong assumption: that AI’s job in a writing pipeline is writing. It isn’t. You just have to write the correct job description.

The job description

Hire the AI for three things — research, structure, and hard questions. Never voice.

The junior-staffer frame is precise, not cute. A good junior staffer doesn’t write the partner’s opinion. They gather the background, propose an outline, and — if they’re brave — ask the partner the questions that expose the gaps: what do you mean by this? where’s the example? what would a skeptic say? The partner’s answers, spoken in the partner’s voice, become the substance. The staffer organized the thinking; the partner did the thinking.

That’s the division of labor, and voice is the bright line. The moment the machine starts producing your sentences, you’ve handed over the only part that was worth anything — the part clients pay for, the part search now rewards, the part Letter 01 calls the asset. Everything on the other side of the line, though — the gathering, the structuring, the interrogating — is work that used to require either hours you don’t have or an editor you can’t hire. Now it’s twenty dollars a month.

The interrogation

Here’s the method, applied to the block-stack you built in Letter 03.

You point the AI at the stack — the question, what most people get wrong, what you actually do — with a prompt that assigns the editor’s job, not the writer’s: find the gaps, propose a sharper order, and ask me the three hardest questions a skeptical reader would ask. (The Vault ships with my exact synthesis prompts; this is prompt #2.)

What comes back is not a draft. It’s better than a draft. It’s the thing every professional writer pays an editor for: “You claim SOPs fail because of ownership, but you never say who should own them — who?” “The middle assumes the reader knows what a runbook is — do they?” “A skeptic would say this only works for big teams. Does it?”

Those questions do something a blank page never does: they make the knowledge in your head retrievable. You’ve answered exactly these kinds of questions a hundred times — on calls, in meetings, under pressure. Being asked is your native mode. Every expert who “can’t write” can answer a pointed question about their own field instantly, fluently, with an example attached. The interrogation converts writing — the thing you’re bad at — into answering, the thing you’re a professional at.

Your answers ARE the writing

Now the part nobody tells you, which is the entire trick:

Answer the questions in your own words — typed fast, or dictated into your phone while you walk. Don’t compose. Just answer, the way you would on a client call. Then paste the answers into the stack.

Look at what happened. You never “wrote.” You talked about work you know cold, under good questioning, and a transcript appeared — in your voice, with your examples, carrying your actual judgment. That transcript, dropped into the stack’s structure, is ninety percent of a piece. The final pass (next letter) is minutes, not evenings.

This is why the voice line matters so much more than a style rule. Voice produced this way can’t be faked and can’t be commoditized — it’s the residue of real expertise answering real pressure. A thousand people can type the same prompt. Nobody else can give your answers.

The test for every AI use, forever

The job description generalizes into a test you can apply to any AI tool, any workflow, any shiny new capability:

Does this use of AI extract more of my thinking, or replace it?

Extraction compounds — every interrogation mines material you didn’t know you had, and the asset that results is more you, not less. Replacement erodes — every generated piece trains your audience that your byline means nothing. Same technology, opposite businesses. The experts who get this right over the next few years are going to be very hard to compete with, because they’ll be publishing at volume and it will all be real.

Do this today

Fifteen minutes:

1. Point prompt #2 (“Interrogate this draft”) at your block-stack from Letter 03. The prompts are in the free Vault if you don’t have them.
2. Take its three hardest questions. Answer them out loud or fast-typed — client-call mode, not author mode.
3. Paste the answers into the stack.

Read the note back. That’s nearly a piece, and you never faced a blank page. Tomorrow we ship it — imperfectly, on purpose.

If what you’re seeing in these letters is a pipeline you want built for you — the vault, the AI staff wired in, the prompts tuned to your field — the Install is the door.

— Frank