Write a Blog Post With ChatGPT, Not Like ChatGPT

Most people use ChatGPT for blog posts the same way: paste in a title, ask for 1,500 words, lightly edit whatever comes back. It's fast, and the result is unmistakable: smooth, confident, weightless, and full of sentences that could have been written about anything.

I found the fix comes from a different division of labour, not a better prompt. This tutorial walks through the workflow that keeps the model on research, structure and criticism, and keeps me on the sentences.

Before you start

A free ChatGPT account covers everything below. The paid tier mainly buys longer context and fewer interruptions mid-session, which matters once drafts get long, though I don't find it required for this workflow.

I also bring one thing the model cannot supply: something I know. A number I measured, a mistake I made, a comparison I ran. Without one, the article turns out generic no matter which tool writes it.

Step 1: Give it your raw material, not your title

I open a new chat and skip typing the headline first. Instead, I dump the unstructured material I already have: my notes, the bullet points from a call, the half-formed argument in my head, the three links I meant to read.

Then I ask for a diagnosis rather than a draft:

Here are my raw notes for an article. Do not write anything yet. Tell me what the actual argument in here is, what is missing, and which two or three points a reader would find genuinely new.

I found this anchors the whole session in my material. Every subsequent answer gets shaped by what I brought, rather than the average of everything the model has read about the topic.

Step 2: Outline in a separate turn, then argue with it

Once I agree on the argument, I ask for structure, staying specific about the shape I want:

Outline this as six H2 sections. For each one, write the section heading and a single sentence describing the job that section does for the reader. No prose.

I get something competent and slightly boring, which is what I expect. Then I push back on it, in the same chat:

Sections 2 and 4 overlap. Section 5 is the interesting one and it is buried. Reorder so the strongest claim lands by section 2, and cut anything that is only there for completeness.

The argument is the point. An outline I've pushed back on twice feels structurally mine in a way a first-pass outline rarely does.

Step 3: Draft the sections yourself, badly and fast

This step decides whether the article sounds human, and it's the one most people skip.

I take the outline into my own editor and write each section from my notes, at speed, without editing. Bad sentences are fine. Fragments are fine. I'm producing my voice on the topic, not a draft, in a form the model can work with instead of replace.

If a section stalls, I don't hand the whole thing over. I ask a narrower question:

I am trying to explain why per-seat pricing punishes small teams and I keep going in circles. Give me three different angles of attack, one sentence each.

Then I write the section myself from whichever angle I like.

Step 4: Use it as an editor, one pass at a time

Now the model earns its subscription. I paste my rough draft back in and run separate, single-purpose passes. One instruction per turn: combined instructions produce mush.

I run them in this order:

1. Structure. "Where does this lose the thread? Point at paragraphs, do not rewrite them."
2. Clarity. "Mark every sentence a reader would have to read twice. Do not fix them, just list them."
3. Cut. "Which 20% of this could be deleted without losing anything? Quote the openings of what should go."

All three ask the model to point, not rewrite. The moment I let it rewrite a paragraph, that paragraph stops being mine and starts sounding like everyone else's. I fix what it points at, in my own words.

Step 5: The hostile read

The last pass improves the article more than any other, and almost nobody runs it:

Read this as a sceptical expert who disagrees with the main claim. What are the three strongest objections? Where am I overclaiming? What have I stated as fact that I should be sourcing?

Then I answer the objections in the text. Articles that survive a hostile read earn links, because they've already done the work the reader was about to do.

The Real Cost in Time

For me, this workflow took roughly the same time as writing it manually, maybe 10% less. The pitch for AI writing usually promises a 10x speedup, and this workflow doesn't deliver that.

Floor quality improves instead: I tighten the structure, stress-test the argument, and handle the obvious objections before the piece ships. I'm trading a small time saving for a better piece.

If you need volume instead, with a consistent brand voice across a team, that's a different problem and a different tool:

Jasper AI Review

See that review for whether the premium is justified.

Where This Fits in the Wider AI Writing Toolkit

This workflow sits in the AI Writing Tools section. Best ChatGPT Alternatives covers the assistants that do the same job, the Google Gemini Review covers the closest free one, and the Jasper Review covers the paid option aimed at teams producing this at volume.

Next Steps

Try the same five stages on your next post, but skip step 4 the first time. Draft, then go straight to the hostile read. Most writers find that pass alone changes more than all the editing passes combined.

For the research stage before any of this, see How to Build a Content Brief in Surfer SEO: the brief is what step 1 above should start from. For what ChatGPT's own plans include, see my ChatGPT Review. For the rest of my AI Writing Tools coverage, see the full hub.