A writing workflow with AI that does not produce filler

1. 1

   1. Research and facts

   You collect what is true: the results page, the documentation, your own data. Nothing is written yet.

2. 2

   2. The brief

   You write one page of instruction: the search, the angle, and a block of facts the draft may not leave.

3. 3

   3. The draft

   The model writes from the brief and nothing else. This is the only step it owns.

4. 4

   4. The human pass

   You check every figure, add what only you know, and cut the register.

5. 5

   5. The read-back

   A checklist or a script refuses what slipped through: invented numbers, banned words, dead links, thin pages.

The model will fill any gap you leave with something plausible. So the first step is to leave no gaps on the things that matter: figures, dates, policies, product behaviour.

• The results page for the query: the ten titles, the People also ask questions, the pages an AI overview cites if one opens.
• The primary documentation for anything you will state as a rule, opened and read, not remembered.
• Your own data: what support is asked every week, what the product does and refuses to do, the figures from your analytics.
• The limits: what you do not know, and will therefore not claim.

This step takes longer than the drafting and it is the one that decides whether the page is worth anything. A page whose facts are all second hand can be written by anybody, including the ten sites already ranking.

A brief is not a topic. It is an instruction specific enough that two different writers would produce the same page. These are the fields worth having, and the last one is the one people skip.

A short brief for a real page looks like this.

SEARCH: “how to invoice a client abroad”, also “international invoice”
H1: How to invoice a client in another country
ANGLE: The three things that actually go wrong (currency on the invoice,
  VAT wording, late payment in another currency). A worked example with
  one fictional invoice. No legal advice; say when to ask an accountant.
READER: A solo founder who bills a handful of clients and has no finance team.
FACTS YOU MAY USE:
- our product converts at the rate on the issue date, not the payment date
- VAT wording differs by country; we do not give tax advice
- the two support questions we get every week are X and Y
- nothing else. No statistics, no percentages, no third-party prices.

The last line of the facts block does most of the work. Without it, a draft arrives carrying an invented adoption statistic and a competitor's price from two years ago.

Hand over the brief and the constraints together: the length, the register, the shapes you want (a table here, steps there), and what to avoid. Ask for the sections you named, not for a page about the topic.

• Draft section by section when the page is long. A whole page in one go drifts in the middle.
• Give the shape you want in the instruction: “a table with these three columns”, “five numbered steps, one sentence each”.
• Forbid the register explicitly. The words you never want are a list you can paste every time.
• Ask it to leave a marker where it wants a fact you did not supply, rather than filling the gap.

That last instruction changes the output more than anything else. A draft with three markers saying a figure is needed is a useful draft; a draft with three invented figures is a trap.

1. 1

   Check every number against its source

   Go through the page and put a finger on each figure. If you cannot name where it came from within ten seconds, cut it. This is not optional and it is where the real risk sits.

2. 2

   Add what only you know

   The mistake your customers make, the screenshot of your own screen, the thing you tried that did not work. One paragraph here is worth more than the rest of the page.

3. 3

   Cut the register

   Remove the throat-clearing opening, the summary paragraph that repeats the section above it, and every adjective that carries no information.

4. 4

   Fix the claims

   Turn “studies show” into the study, or into nothing. Turn a hedge into a decision. A page that will not commit is a page nobody trusts.

5. 5

   Read it aloud once

   The sentences that were assembled rather than written are audible. You will hear them before you see them.

The human pass is tired by the fourth page. A read-back is mechanical and does not get tired. These are the checks worth automating, and they are the ones this library runs before a page is published.

• Any number on the page that was not in the brief, flagged for a human to confirm or cut.
• The banned register words, as a straight word list.
• Dashes and exclamation marks, which a model reaches for and a house style may forbid.
• The title over its character budget, and the description outside its range.
• Links that point at pages which do not exist.
• Pages under a word count that says nothing was really covered.
• Pages nothing else links to, which nobody will find.

A spreadsheet formula and a careful proofreader catch most of this. The point is that it happens every time, on a rule, rather than when somebody remembers.

Without a brief, a model writes the average of what it has read on the subject. The result is a page where every sentence sounds right and the page as a whole contains nothing: no figure you can act on, no example, no decision.

That page is not just a waste of an afternoon. It is the pattern Google's scaled content guidance describes, published at any volume, and it is exactly what the ten pages already ranking are made of. You cannot win that query with more of the same.

The workflow above is a way of making sure the thing you own, your facts and your judgement, ends up on the page, and that the model only does the typing.

For a guide of fifteen hundred words: the research and the facts take the longest, the brief takes twenty minutes, the draft comes back in minutes, and the human pass takes about as long as the brief.

The saving is not on the first page. It is on the tenth, because the facts blocks are reusable: the documentation you read once for one page is the facts block of the next four in the same family.

We make Porteur, which does the research half for a whole site: the searches, the rivals on them and the gaps, written as briefs.