Writing with AI: the professional workflow

Which part of the job is the model and which is you

Every text written with AI passes through five stations, whether you name them or not: brief, draft, fact pass, voice pass, publish. Where it goes wrong is almost always the third one, and the reason is not technical: nobody owns it. The model assumes you will check, you assume that fluent means correct, and neither assumption is written down anywhere a person would read it.

So the first thing a professional workflow does is not produce a better output. It gives every station an owner. Two of them never leave human hands at any price: deciding what is worth saying, and deciding what is true. Two others, assembling and rearranging, are what the model is genuinely good at. The voice station is shared, and most of your time goes there.

That division has one consequence that runs against instinct: the cleaner the draft, the harder the fact pass. A text written confidently and without seams does not show you where to look, so you have to read all of it at one uniform level of suspicion, which in practice means reading none of it. That is why at the second station we deliberately ask the model to leave its seams showing.

One boundary is worth stating up front: this workflow is about assistance, not bulk production. We looked at the difference against the text of Google's own policies in the lesson on AI content and SEO, and the position here is the same one: assisted and verified gets published and stays, bulk and unchecked gets published and dies.

The brief: the one thing the model cannot write for you

A brief is a page, not a sentence, and every decision in the text is made there. If the brief is empty the model guesses, and you mistake editing its guesses for writing. That is the worst case, because it costs time and it feels like work.

A brief that earns its keep holds six things:

• Who the reader is and what they already know.
• What they should be able to do after reading.
• What you can actually deliver. Anything outside that is not a promise.
• The terms you use, and the terms you refuse to use.
• The sources the text is allowed to take numbers and claims from. Only those.
• The thing only you know, which is in no source.

The fifth item does more work than the rest and gets written least often. With a closed source list the fact pass turns from a judgement job into a matching job: every claim is either inside that pack or it is not. With an open list the model writes from memory and you have to verify every claim from scratch.

The sixth is what separates your text from every other text on the topic, and it is the only line no model can fill. If it stays empty, the piece is not ready to be written and the problem is somewhere else.

There is a simple test. Read the second item out loud. If you cannot say in one sentence what the reader can do afterwards, no model solves that, and asking for a draft is only postponing the decision.

The draft pass: get a text you can actually edit

Two decisions get made at this station: which class of model, and which constraints.

The class follows the kind of work, not the budget. Assembling, rearranging and compressing something you already have is mechanical, and a fast cheap model of the Flash class handles it. Where judgement is required, like choosing an angle or spotting where an argument gives way, a frontier model earns its cost. Which model sits in which class today is written up with sources in the AI section of this site, and we do not print a version name here, because any name we print will be wrong in six months.

The constraints are a different matter, and the ones that work are negative. Three that genuinely do: do not write any number that is not in the brief; mark anything you added yourself in brackets; and do not use these words, followed by your own list.

The second constraint is what makes the next pass possible, and almost nobody uses it. A draft with its seams marked is ugly, and that is exactly why it is useful: you know where to look. A clean seamless draft hands you an invisible checklist that you will never work through.

There is one habit that visibly improves the output and is written down almost nowhere: instead of asking for the text, ask for the structure first and wait for approval. A wrong structure is visible in thirty seconds and invisible inside two thousand words. Every time we have skipped that step, a second draft has been needed.

The fact pass: where did each number come from

This is the station where AI writing usually fails, and it fails quietly. The model writes a number, a date, a quotation and even the name of a source with exactly the confidence it uses for true things, because statistically that is what usually appears in that position. A text with one invented number is a text where, once somebody catches it, no other sentence carries weight either. The damage is not proportional to the size of the error.

The rule is simple and it is written into our own content law: a number without a source does not get written. If a number would strengthen the piece and you do not have it, you leave a marked placeholder and declare it at handover; something that looks like a number does not take its place. The difference between those two moves is zero on the page and total in the outcome.

Four things the model invents without warning, and this pass hunts all four: a number, a date, a quotation, and a source. The last is the most dangerous, because a fabricated source looks more respectable than a fabricated number and fewer people click it.

The method is mechanical and should stay mechanical. You pull every checkable claim out of the text, write next to it which source in the brief pack it came from, and quote the exact sentence from that source. A claim with no sentence behind it is deleted or kept with a mark. This takes patience and no skill, which is precisely why it is the best candidate for handing to a machine. The recipe is further down.

The voice pass: where the text starts sounding like you

Voice is not adjectives. Voice is what you decline to say, where you admit you do not know, and what rhythm your sentences have. The model has none of those three of its own, and if you ask it to make the text "more human" it adds decoration: more adjectives, more similes, the same skeleton underneath.

Three specific moves happen at this station. First, break up the rhythm, because uniform sentence length is the loudest signal of machine text. Second, delete the opening that could sit on any article about anything. Third, put in at least one first-hand thing only you know. All three are worked through, with method, in the next lesson of this path, on humanising AI text.

There is a boundary here that sounds rigid and is not: you are not allowed to add new claims during the voice pass. If you realise mid-edit that a new claim is needed, you go back to the fact pass and bring its source. The reason is experimental: a claim added during the voice pass never gets checked, because every checking station finished before it.

Publishing: what must be settled before you press the button

One thing: the name of the person standing behind the text. The editorial policy of this site puts it in a sentence: no content is published without review and approval by a human specialist, and responsibility for accuracy and final judgement sits with the human author, not the machine. It sounds strict and in practice it is the easiest part of the job, because it asks only one question: if somebody objects to a sentence in this text tomorrow, who answers.

Whether to declare the use of the tool is a second question, and our answer is this: more important than declaring the tool is that a specific person stands behind the accuracy of the text. The practical test is simple. If a reader would feel misled on learning something about how the text was made, write that thing down. If not, declaring it is one more paragraph and nothing else.

And one small thing that works better than it looks: print the date the facts were last checked. A text that says it was checked on a given day lets the reader judge for themselves which part may have gone stale, and that is exactly what an undated machine text never gives anybody.