What this article covers
I am out walking, listening to YouTube, and I hear a stretch of something genuinely well put. I send the link from my phone with a few sentences of voice note, and the computer at home starts analysing it, drafts a short post in my own voice, makes the image cards, and publishes. By the time I get home the work is sitting there waiting for me to confirm.
This article takes that loop apart: which stages I hand over, which single gate has to stay with a human, and the four checks I wrote into it. It also covers the two gaps I found after this run, because that is what the loop actually looks like.
- People with something they have to produce regularly (posts, reports, assets) and who start from scratch every time
- People whose ideas keep arriving when they cannot sit down and work
- People who want to know how far automation can actually go, and where it should not
- The five stages of a complete publishing loop, with a completion condition for each
- Four image checks you can copy directly
- A way to decide which gate you cannot hand over
What is written here reflects August 2026. Phone-to-desktop remote calling is the kind of feature where support conditions, plans and versions are all still changing, so check the vendor's current documentation before you rely on it.
1. How this started
My old approach to that moment was to save the link and deal with it when I got home. In practice, most of the time there was no next step, because other things were waiting at home.
What I do now: send the link from my phone, record a few sentences about the angle I want, and carry on walking.
When I surveyed the room, the people who knew about this and the people who did not were roughly half and half. One participant pointed out that another tool has the same capability, and that is correct. I have used both; the flow described here was run on one of them. How smooth it feels depends on your account plan, the platform version and your own workflow, so both routes are worth trying yourself before deciding. Also, it is no longer just phones. A laptop can drive the desktop machine now too.
2. The five stages of this loop
Phone sends link plus voice
I do not organise it and I do not write a headline. I just say the angle I want. Organising is its job.
Desktop analyses the content
Pulls the content, cross-references what I have written before, drafts a short post in my own voice.
Stops for a human
The copy gate is not automatic. This is the only human gate in the whole line.
Makes the cards and checks itself
Generates and lays out the images, then verifies them against four standards.
Publishes to three platforms, and every platform has to report success or failure.
Of the five stages, only the third belongs to a human. The other four run on their own, and the run stops at stage three to wait for me.
3. Which gate has to stay with a human
I have never handed over the copy gate, and the reason is simple:
That sentence is the boundary of this loop. It can pull content for me, organise the structure, cross-reference what I have written before. But "what I think about this" is the part where handing it over empties the whole thing out. What goes out carries my name.
That day I did send it back to be rewritten: I felt it read too professional, too much like a summary of someone else, not like the way I talk. I only moved on after its second version.
4. The four image checks written into the loop
The image-card stage is fully handed over, because its standard can be stated clearly. I hard-coded four items:
Format
Is it 4:5. Easiest to verify, and the one that goes wrong most often.
Compression and conversion
Has it been compressed, has it been converted to the format it should be.
Text
Has the text on the image drifted. It has to match the script word for word.
Character design
Confirm it is still the same character and the same style.
These four are exactly what I used to look over by hand every single time. They have one thing in common: each can be stated in one sentence that says what "correct" means. What can be stated that clearly can be written into a rule and handed over. What cannot is the kind of gate the previous section says to keep.
You can turn this straight into your own version:
5. This run only half worked
On the day of the talk I showed the actual run of this loop, including the parts where it did not work. And one of the problems I only noticed while I was teaching.
The problem was that my rule said "check twice", so anything from the third image onward was never covered. That is a hole in the rule, not laziness on its part.
The second gap is the one worth talking about: it did not publish successfully, and it also did not tell me it had not. The whole line quietly skipped two thirds of the publishing.
6. The fix: turn what was missed into a checkpoint
What I did afterwards was simple, two sentences:
- "Check twice" became "every single image has to be checked". The original wording treated a count as a scope, and that was my own mistake.
- "Publish" became "each of the three platforms reports success or failure separately". No success report means it is not done, and the run does not count as finished.
For the two platforms that did not go out, I went back and had it publish them. Before doing that, go and look at those two platforms to check the post really is not there, or you end up publishing the same thing twice. What actually needed fixing was the rule, not that one run.
With those two lines added, this loop moves forward one notch next time. Loop engineering is not something you write once. It is adding a checkpoint each time you find a missed check.
7. If one route is blocked, switch: the other loop for viewpoint reports
I run the same approach on another line for heavier material: someone explaining an AI paper, or an interview with a well-known figure. I turn those into viewpoint reports for my own internal reference, not into posts.
The key to this line is the switching rule:
Pull the captions directly and carry on to analysis.
Put it through something like NotebookLM to produce text, download it, and rejoin the original flow. If what you upload involves client or company internal material, confirm that service's policy and your own organisation's rules first.
Stop and report which video it was and which step it stuck on. Do not force it through.
The important part about writing a switching rule is that you have to have walked the backup route yourself once first, or it has nothing to follow. I had written up how I put a video in and pull the result back out, so now it just follows that.
8. Why it fills in things I wrote before on its own
One part surprised me on the day too: while writing this piece it went and added material from another topic I had written about earlier, and I had not asked it to.
The reason is the shared knowledge base: my documents, analyses, strategies, plans and skill packages all sit in the same folder, and the AI can read it. So when it writes something new, it can find what I said before, and it knows the two have to line up.
Put the other way round: if your material is scattered across separate chat windows, this does not happen, because it has no idea any of it exists.
How you can start
Pick something you produce two or three times a week, map out its stages, and mark which stage going wrong is a "format mistake" and which is a "who I am mistake". The latter is the gate you cannot hand over.
Use the format in section four to write down the three to five items you look at every time, and remember to write "every one of them" instead of a count.
Think about the step where you get stuck most often, walk the backup route yourself once, then write it as "if this route is blocked take that one, and if both are blocked stop and report".
Not having a desktop tool is fine too. Sections three, four and six work just as well in a browser. The difference is that you press start yourself, instead of it running while you are out.
Three related articles
- Why Does the AI Keep Forgetting What I Told It?: From "I'll Remember That" to Letting It Run Itself: once the rule is written, how to confirm it really was remembered and really gets triggered.
- Will Someone Else's Skill Package Mess Up My AI?: Three Things I Do Before Copying One: what to look at before installing a workflow somebody else wrote.
- The AI Revised It Ten Times and It Is Still Wrong. Now What?: Cleaning the Data Is Only the First Layer: the full order of operations when you need an AI to check a long document.