My actual human-in-the-loop workflow, step by step

I have written about why I keep a human in the loop and it was mostly argument. Several people have since asked the more useful question, which is what that actually looks like on a Tuesday. So this is the mechanics, in order, including what happens when a step fails.

The rule underneath all of it is simple and I apply it without exception: the machine does the grinding, a person makes every call that has a consequence. Everything below is that rule turned into steps.

Step 1: decide the page exists. Me.

Nothing gets generated until I can say what search term it owns, who it is for, and what it is supposed to make somebody do. If I cannot say those three things, the page does not get written, and quite often that is the right outcome - the majority of the ideas that occur to me do not survive this step.

This is also where I check nothing already owns that term. For my own site I keep a sheet with every URL and the one term it owns. It is boring and it is the single most valuable document I have.

The machine does nothing here. It cannot: it does not know what is already on the site, what I want to be known for, or which of two ideas is worth the afternoon.

Step 2: assemble the context. Me, then the machine.

I paste in a document of my own positions and phrasing - what I think about the thing, how I explain it, the numbers I am allowed to use, and a hard instruction that if a fact is not in the document it gets flagged rather than filled in.

For anything touching the study data, the context includes the actual figures, because that is the one place an invented number would do real damage. Every figure I publish traces back to a document that is the single authority for it, and the drafting step is not allowed to produce a number that is not in there.

Step 3: structure. Me.

I write the headings myself. Not the prose - the headings, in order, plus a line under each saying what it has to do. This is the step people skip and it is the one that determines whether the piece is any good, because structure is argument. A model given a good structure writes a decent draft; a model given a topic writes an encyclopedia entry.

Step 4: draft. The machine.

This is the only step that is genuinely automated, and it is the fastest one. Headings plus context plus a brief, out comes a draft. It takes a couple of minutes and it removes the blank page, which is the thing that has kept half-finished websites half-finished for years.

The draft is never the thing that gets published. I would put it at sixty per cent of the way there on a good day.

Step 5: check the claims. Me.

Every number, date, name and claim gets read against the source. Not skimmed - read, one at a time. This is the slowest step and I have never found a way to shorten it that I trust, because the whole failure mode is that a wrong fact reads exactly as smoothly as a right one.

If a draft has invented anything, I do not fix it and move on. I go back to step 2 and add the missing fact to the context document, because a gap that produced one invention will produce another. The order matters in editing AI output.

Step 6: put the specifics in. Me.

The draft says "significant improvement". I know the actual figure. The draft says "many businesses". I know which ones and what happened. This pass is where a generic piece becomes one that could only have come from me, and it is almost entirely additive.

Step 7: rewrite the opening. Me.

The first two paragraphs get written by hand, from nothing. Every time. That is where a reader decides whether there is a person here, and it is the part a model writes most averagely - which is covered from the reader's side in readers can tell.

Step 8: links. Me, with the machine typing.

I decide what this page links to and what links to it, and with what anchor text. If I want the drafting to help, I hand over the actual list of URLs and tell it to use only those - it cannot invent a link to a page that does not exist if it only has real ones to choose from.

Then I go and add the inbound links to the older pages. This is the step that gets skipped most often, including by me: I published a set of nine posts recently where the anchor guide linked to none of its own spokes for a fortnight, which made every one of them a dead end from the top of the cluster.

Step 9: read it out loud. Me.

If a sentence is hard to say it gets cut. This catches the balanced, rhythmic, says-nothing paragraphs better than any amount of staring at the screen.

What gets thrown away

Roughly one draft in five does not survive. Usually because step 5 turns up something I cannot substantiate and the piece was resting on it, or because reading it out loud in step 9 reveals that I did not actually have a point. Both are cheap failures at this stage and expensive ones after publishing.

The thing I would want you to take from that is that the throw-away rate has not gone down since I started drafting this way. Generation made the writing cheaper. It did not make my ideas better, and pretending otherwise is how sites fill up with pages nobody needed.

How this post was made

Exactly as above. I decided it should exist because people kept asking the mechanics question after reading the principle. I wrote the nine headings. The draft came back in about two minutes. Step 5 was short because there is little to fact-check here, step 6 added the fortnight-of-dead-ends admission, and I rewrote the opening twice because the first version started with a definition, which is the tell I complain about in other people's writing.

I am comfortable telling you to use these tools because I use them, on this site, on exactly these terms. The reasoning is in why I keep a human in the loop, the day-to-day in how I actually use AI for SEO, and the structure this all sits inside is using AI for your website content without wrecking it.

What I do not let it do at all

There are steps I would not hand over however good the tooling gets, and they are not the hard ones. They are the consequential ones.

• Decide what to write. That is a judgement about what I want to be known for, and it is most of the value.
• Decide which page owns which search term. Get this wrong and two of my own pages compete forever.
• State a figure. Every number I publish comes from a document that is the authority for it, and the drafting step is not permitted to produce one that is not in there.
• Choose anchor text or where a link goes. That is a decision about what I want to be considered important, which is precisely the thing a tool with no view of my priorities cannot make - and the reason I am sceptical of tools that automate it, set out in the anchor text red flag.
• Decide a draft is finished.

Notice that none of those are writing tasks. That is the whole point: the writing was never the part that needed judgement, which is why handing over the writing is fine and handing over the rest is not.

Where the time actually goes

Roughly, for a post like this one:

• Steps 1 to 3, the deciding and the structure: about 40% of the total.
• Step 4, the generation: under 5%, and it is the only part that got faster.
• Steps 5 to 7, checking and specifics and the opening: about 45%.
• Steps 8 and 9, links and reading aloud: the rest.

So about eighty-five per cent of the work sits either side of the bit that was automated. That ratio is the honest answer to "how much time does AI save on content", and it is a lot less than the pitch implies - though the part it removes, staring at an empty page, is the part that used to stop things happening at all.

If you want to copy it

The steps that matter most, if you only adopt some:

1. Step 1, because it prevents the pages that should not exist.
2. Step 5, because it prevents the claim that costs you.
3. Step 7, because it is sixty seconds and it does more for conversion than the rest combined.

Steps 3, 6 and 8 are what separate a good piece from an adequate one. Step 9 is free. Step 4, the one everybody focuses on, is the least of it.