Most people use an AI writing tool the same way: type a topic, press generate, read the output, decide it is bland, and then either publish it anyway or go back to writing by hand. Both outcomes waste the tool.
The drafts that actually get published come out of a different process, and generation is the short part of it. What surrounds it — a specific brief before, a structural edit after — is what separates a post worth your byline from filler. This is that process, in the order you should run it.
Step 1: Decide what the post has to do
Before you open a generator, answer three questions in one sentence each. Who is reading this, what do they already know, and what should be different after they read it? A post for someone still choosing between tools looks nothing like one for someone who has already bought and is stuck.
This is not ceremony. Every detail you leave unspecified gets filled in with the most average possible answer, because that is what a language model does with an unspecified detail. A vague brief is a request for generic writing.
Step 2: Write a brief, not a topic
A topic is “email marketing”. A brief is “a walkthrough of setting up a three-email welcome sequence for a small ecommerce store, for someone who has a list but has never automated anything, focused on what to send rather than which tool to use”.
The second one constrains the output four ways, and each constraint removes a category of generic sentence the model would otherwise reach for.
- Scope — what the post covers, and explicitly what it does not.
- Audience — their role, plus the one thing they already know so the draft can skip it.
- Angle — the specific claim or approach the post argues for.
- Exclusions — the obvious sections you do not want, such as a history of the topic.
Step 3: Generate the outline before the draft
Produce the structure first and read it critically. An outline takes seconds to generate and seconds to judge, whereas a full draft built on a bad structure costs you an entire edit to discover the same problem.
Look for three failure patterns: sections that restate each other in different words, a section that should have been the whole post, and the throat-clearing opener that adds nothing. Cut all three before you generate a word of prose.
Step 4: Generate the draft in sections
Generating 1,500 words in one pass gives you one thing to accept or reject. Generating section by section gives you control over pacing, and lets you feed in specifics — a number, a customer objection, a product constraint — exactly where they belong.
Set the tone once
Pick the tone and length up front and hold them constant across sections. Switching tone midway is the clearest sign a piece was assembled rather than written.
Write the body before the intro
You cannot introduce an argument you have not made yet. Draft the middle, then open with a sentence that earns the rest of it.
Keep the outtakes
A discarded paragraph often holds the one sharp sentence its replacement lacks. Park them in a scratch file until the post is done.
Step 5: Restructure ruthlessly
Generated drafts have a characteristic shape: every section roughly the same length, every paragraph three sentences, every sentence between fifteen and twenty words. It is readable and completely flat — and readers feel the flatness even when they cannot name it.
Break the pattern on purpose. Put a four-word sentence after a long one. Let one section run twice the length of the others because it earns the room. Delete any paragraph that exists only to get you to the next one.
| Tell | What it looks like | Fix |
|---|---|---|
| TellUniform rhythm | What it looks likeEvery sentence 15-20 words | FixCut three sentences in half; merge two others |
| TellHedged claims | What it looks like“Can often help improve” | FixState it: “cuts drafting time by about half” |
| TellSummary paragraphs | What it looks likeA section restating the one above it | FixDelete it entirely |
| TellListicle padding | What it looks likeTen items where four carry the value | FixKeep four and expand each |
Step 6: Add what the model could not know
This is the step that decides whether the post is worth publishing at all. A language model can produce a competent explanation of almost anything. It cannot produce your numbers, your customer’s exact wording, the thing you tried that failed, or the screenshot from last Tuesday.
- One concrete number from your own data, with the context that makes it mean something.
- One objection a real customer raised, answered honestly.
- One thing that did not work, and why you think it failed.
- One specific named example in place of a hypothetical one.
Four insertions is usually enough to turn a competent draft into something only you could have published.
Step 7: Check before you publish
Run the finished post through three quick checks. Readability: is the grade level where your audience actually reads? Repetition: has one phrase crept in twenty times? Headline: does the title survive truncation in a search result and still promise something specific?
Then verify every factual claim. AI drafts state wrong things with exactly the confidence they state right ones — statistics, dates, product capabilities and quotes are the four categories that most often need correcting.
How long this actually takes
For a 1,200-word post: ten minutes on the brief and outline, two minutes generating, thirty to forty minutes editing and adding specifics. Call it an hour, against three or four hours writing from scratch.
The saving is real, but notice where it comes from. Generation was never the expensive part of writing — the empty page was. AI removes the empty page. It does not remove the thinking, and any workflow that pretends otherwise produces posts that read exactly like it.