The standard AI writing workflow is a conversation with one model. You ask, it drafts, you critique, it revises. Six rounds later you have something acceptable and a vague sense that you did most of the work.
There is a better shape, and it comes from a simple observation: the second answer from the same model is usually a variation on the first. The first answer from a different model is usually genuinely different.
Iterating with one model narrows. Asking several widens. For the parts of writing where you are stuck — angle, structure, opening — widening is what you need.
Here is the workflow.
Why one model narrows
Every model has a house style. Not just tone — structural habits. Preferred paragraph rhythm, characteristic transitions, a default way of organising an argument, a set of framings it reaches for first.
Ask the same model to “try again” and it explores the neighbourhood of its first answer. That is useful for refinement and useless for escape. If the first draft was aimed at the wrong angle entirely, ten revisions produce ten well-polished versions of the wrong angle.
Different models were trained differently and have different habits. Asking three of them the same question surfaces three genuinely distinct approaches, and the gap between them is where the interesting choice lives.
The five-stage workflow
Stage 1 — Angle (multi-model, wide)
Send the same brief to three or four models, unchanged. No refinement, no follow-up questions. Ask each for five angles on the topic with a one-line rationale each.
You will get fifteen to twenty angles, of which perhaps twelve are obvious and overlapping. The value is in the three or four that only one model produced. Those are the ones worth considering, because they represent a framing you would not have reached alone.
Pick the angle yourself. Do not ask a model which is best; it will pick the most conventional one, because conventional is what “good” looks like statistically.
This stage requires running one prompt across several models quickly. Comparing outputs side by side is the whole point, and side-by-side model comparison is faster than opening four tabs and scrolling between them. Multi-model platforms such as Perspective AI let you send one prompt and read the answers next to each other in a single thread, which turns this stage from a chore into a two-minute step.
Stage 2 — Structure (one model, deep)
Now narrow. Take the chosen angle to a single model with a strong long-form reasoning profile and work the outline properly: sections, the argument each one carries, the transitions between them.
This stage is conversational. Push back, ask for alternatives to specific sections, question the ordering. One model, several turns — the opposite of stage one.
Stage 3 — Draft (one model, your preferred voice)
Draft with whichever model you have found produces prose closest to your voice. This is genuinely subjective and there is no benchmark for it. Test three models on a piece you have already written well, and use the one whose output needs the least rewriting.
Give it the outline, your voice notes, and one or two paragraphs of your own writing as a reference. Reference text does far more for voice matching than any amount of adjectival instruction.
Stage 4 — Adversarial review (a different model)
This is the stage most people skip and the one with the highest return.
Take the draft to a different model from the one that wrote it and ask it to attack:
- Which claims are unsupported?
- Where does the argument skip a step?
- What would a sceptical reader object to?
- Which paragraph is doing the least work?
A model reviewing its own output is a poor critic — it produced that text because it found it good. A different model has no such attachment and will find real problems.
Stage 5 — Edit (you, alone)
No model. Read it aloud, cut 15%, fix the rhythm, replace the phrases that sound like everyone else.
This stage is not optional and it is not delegable. It is what makes the piece yours rather than competent.
What this workflow is actually good at
Escaping a stuck opening. Three models, three openings, pick the one that surprises you. Fastest cure for a blank page there is.
Finding the non-obvious angle. The value is in the outlier suggestion, not the consensus one. Consensus is what everyone else already published.
Catching your own blind spots. The adversarial review stage finds the assumption you did not notice you were making.
Speed on structured formats. Comparison pieces, how-to guides, and listicles have predictable shapes that benefit enormously from parallel drafting.
What it is bad at
Anything requiring your specific experience. A model cannot write the paragraph about what went wrong at your company in 2024. That paragraph is the reason anyone reads the piece.
Voice, without heavy editing. Multi-model drafting produces competent prose. Competent is the baseline, not the goal, and the gap between competent and good is entirely stage five.
Facts. Every model will produce plausible statistics with no source. Every number you keep must be verified independently. This is not a flaw to be prompted away — it is how the technology works.
Short pieces. For anything under 500 words the workflow overhead exceeds the benefit. Write it.
The practical setup
Three things make this workable day to day.
Fast switching. If comparing models means logging into different accounts, you will do it twice and then stop. The switch has to be a dropdown.
Pooled usage. Under per-vendor subscriptions, using four models means four bills, and stage one becomes economically absurd. Pooled credits across models is what makes parallel prompting affordable enough to become a habit.
A saved brief template. Reuse the same structured brief — audience, angle, length, voice notes, what to avoid — across models. Comparing outputs is only meaningful when the input was identical.
A worked example
Brief: “1,500-word piece for freelance designers on pricing their work.”
Stage 1 across four models produced fifteen angles. Twelve were variations of “value-based pricing” and “know your worth”. Three were not: pricing as a client-filtering mechanism, the psychology of the first number said aloud, and what changes when you raise prices on existing clients rather than new ones. Only the third appeared from a single model, and it was the most interesting.
Stage 2 turned that into six sections with one model over about ten minutes of back-and-forth.
Stage 3 drafted it in roughly the right voice with two reference paragraphs supplied.
Stage 4 with a different model flagged that section four asserted a claim about client retention with nothing behind it. It was right, and the claim came out.
Stage 5 cut 300 words, rewrote the opening entirely, and added the one anecdote that made it worth publishing.
Total: about 90 minutes. Most of the improvement over a single-model process came from stages one and four — the two that use more than one model.
Frequently asked questions
Does using multiple AI models actually improve writing quality? It improves the range of options at the ideation stage and the quality of critique at review. It does not improve prose quality directly — that comes from editing.
How many models should I compare? Three or four at the angle stage. Beyond that, returns drop sharply and you are just reading more of the same.
Which model is best for writing? There is no defensible general answer, because voice preference is subjective. Test candidates on a piece you have already written well and pick the one requiring least rewriting.
Is AI-assisted writing penalised by search engines? Search guidance targets low-value content regardless of how it was produced. Original reporting, real expertise, and genuine editing are what distinguish useful pages — production method is not the variable being judged.
Do I need multiple subscriptions for this? Not any more. Multi-model platforms provide several labs’ models under one account with pooled usage, which is what makes parallel prompting economically sensible.
The short version
Widen at the start, narrow in the middle, attack at the end, and edit alone.
One model gives you one perspective, repeated with increasing polish. Several models give you a choice — and choosing well is most of what writing is.
Leave a Reply