The Social Media Hotelier
← Letters

The system passed every check. Then it posted 4 sh*t images.

The Social Media Hotelier8 min read

A few months back, I smugly boarded a plane to Malaga knowing that I had stress-tested my AI-powered content pipeline to within an inch of its life.

That test ran 3 posts a day across 6 platforms for 10 days straight. 180 test posts. My quick maths had me convinced that was enough to guarantee the second exercise would take care of itself: 1 post a day on each of the 6 platforms for 16 days, whilst I got on with quality family time.

5 days after we flew to Spain, carousel posts were going out with, frankly, the sh*tiest graphics... despite all the checks and balances I'd put into the process.

This is the story I promised about my first extended attempt to automate the 10-80-10 Protocol I've been banging on about all year.

The 4 of us were driving down from the Sierra de las Nieves mountains to coastal Cadiz when my phone finally reconnected itself with the rest of the world. I cleared through the dozens of pings it had sent me and thought, "let's have a quick peek at how my automated content has been performing".

The management system I'd built to handle all the publishing had marked everything as checked, scheduled, rechecked and posted cleanly.

Excelente.

Now to check the actual performance of the posts. Halfway down a mountainside, I squinted at my small screen. What I saw on my Instagram profile was a mess of half-rendered images, pixelated graphics and carousel decks shuffled up so they made no sense.

No bueno.

The posts went out with every graphic blockier than Minecraft. The files had been shrunk before publishing. Post your content at full resolution, 1080x1350px, every time.

Before I tell you what I did about it, I should be clear about what I was experimenting with in the first place.

I wasn't trying to funnel bookings (I closed my guesthouse in Oxford at the back end of 2024). I coach people like you on how to use social media to drive potential guests to booking pages. So August's efforts weren't a growth experiment. They were to prove, or disprove, a theory I've been working on. If I automated as much of my 10-80-10 Protocol as was technically possible, would it hold while I enjoyed this view from a sun-lounger without an internet connection?

The 10-80-10 Protocol

Three parts to this AI-powered content strategy.

The first 10% is Clarity. That's the Commander's Intent. You decide who the content is talking to, what keeps them up at 2am, what success looks like, and what they want to hear from you.

The middle 80% is the making. Drafting, formatting, creating. AI has commoditised this part and it's almost free.

The last 10% is Human Refinement. That's the Command Inspection. You read back what came out and decide whether it's actually yours or beige slop. Apply your own taste and tone, tweak it, approve it. Only then does it ship or schedule.

The first 10%: what I decided

Two decisions mattered, and both of them were mine.

The first was what to actually build. I interrogated the brief before writing a line. Did I want an agent that wrote the copy, or one that only vetted it? Which platforms would prove the system? What platform-specific idiosyncrasies did I need to factor in? I'm going to need accompanying graphics, at least for Instagram, so could it create those too, in a branded style? And where would I inject myself, like a sentry on a gate?

That interrogation piece is the whole point, and skipping it is where most AI projects die. The machine can't infer any of it. You’ve gotta get your brief right. Every hour I spent there saved me ten further down.

The result was a content pipeline that grew in scope. It needed a documented index, 9 approval gates where I stayed in the loop, and a locked structure for every post, tailored for Instagram, Facebook, Threads, Substack, LinkedIn and X.

Of course, I used AI to help me outline the system. That's how I learned about harnesses, graph prompting, loops, and MCP. I also found out how stubborn I'd have to be to keep pushing for the standard I wanted.

The second decision was what to feed it. Because crap inputs lead to crap outputs, I was dead set on the AI producing content that looked, sounded, felt and smelt as close to my manually-created work as possible. If that went well, the final 10% would shrink to almost nothing.

So I threw the kitchen sink at it, alongside the bath, the bath-water and the baby.

  • 332 email newsletters
  • 300+ hours of transcripts from coaching calls
  • 50+ hours of my digital training programmes, and the prep documents and script outlines behind them
  • 32 blog articles
  • 8 transcripts from podcast appearances
  • 1 granularly defined Ideal Client (yeah, that's you)

The last 10%: what I approved

The last 10% needed another human sentry before anything was scheduled. I read what came back, tweaked it, added my taste and approved it for public consumption.

The clearest example is the Critic Loop I built. On the first pass at one piece of content, my Critic scored the draft 96 out of 100. I was elated. The system was finally fine-tuned. Almost on a whim, I told it to drop any flattery, lose any sycophancy and grade again. The same draft scored 72.

We ran the loop until it shipped at an honest 90.3, and I made zero sycophancy a permanent rule throughout the chain of processes I was building.

That's the Command Inspection working: tweaking, rejecting anything beige, adding my own taste until I was happy to approve the output and get it scheduled.

It's also the step in the framework most people skip, because approving feels like a formality... right up until you catch something you'd never have signed off.

I buckled up in my EasyJet seat with the queue approved and loaded. 1 atom of value per day, 6 platforms, same slot every evening, for 16 days straight.

Where it broke

Both 10s held. The Clarity was mine, the Refinement was mine. The break came in the delivery layer that followed all of the above: it happened after I'd signed everything off.

8 August: 1 image went out with a grey block across its lower two-thirds.

9 August: 4 shrunken graphics files, 3 repeating the same text, all somehow passing verification.

I found out why when I got home. The upload path/scheduler was forcing every file to fit under a size limit. The automated clearing check only compared what the system had intended to publish with what was actually published. Of course they matched. The AI couldn't spot the difference. It had no eyes.

What it actually cost to fix

About 2 hours and 15 minutes, maybe 2 and a half hours, spread across two days. That's what it took to remove the broken posts and re-post them properly.

I could access all the original imagery on my phone and tablet. I had copies of every accompanying caption catalogued by date and by platform. Mercifully, WiFi is easy in Cadiz. So I recreated the posts manually, deleted the corrupted ones and republished them by hand.

Then I checked all the images coming down the pipeline. I made sure revised checks were in place and stepped away again.

I handed control back, and the machine finished the fortnight without me. The audit I ran on my return to Oxford found 40 of 44 automated assets clean, with the damage confined to 8 and 9 August.

The process broke once, in one 48-hour window, and the fix wasn't a total rebuild. It was 2 and a half hours, a potentially over-anxious backup plan and a clear checklist.

The numbers, for transparency

Instagram carried the reach: 3,746 views, up 19%. Unique reach 596, up 106%. Profile visits 88, up 69%. Disappointingly, Instagram delivered 1 very lonely link click across the fortnight, and 8 follows against 10 unfollows.

LinkedIn gained 25 followers over the same fortnight, the best growth on any platform, while Instagram went backwards by 2. And the two platforms wanted opposite things. Instagram gave its best numbers to rule-and-list posts, and its worst to the story posts. On LinkedIn, the result flipped. The RAF Selection story that underperformed on Instagram became LinkedIn's best post of the run, at twice its baseline.

Meanwhile 3 of the 6 channels were close to dead air. X produced 110 views across the entire fortnight and 0 likes. Facebook fell 83% against the fortnight before, at 5 to 52 views per post. Substack Notes added a single follower in 16 days.

One more number, because this email is about transparency. I posted 7 times a week where my niche peers posted closer to 4, and my engagement per post was 0.11% against their 1.7%. Volume was never the win. It never is.

This test wasn't designed to optimise my content. It was designed to assess whether the process held. Other than 1 blip, it did.

1 number in particular, though, made me smile. A Story I posted by hand on 13 August while leaving a peña flamenca took 349 views and 271 reach. It was just a shot of a dancer, captioned “Increíble... otra vez.” More than any scheduled feed post all fortnight. The human layer did more than inspect. Sometimes it just delights.

The question

The skill worth building is knowing enough about a job to supervise the work you hand over.

If you don't know enough about the work, you can't hope to hand it off to a VA, a social media manager, or an AI. Whether you're handing over to a person or a machine, you have to understand what you want from them and communicate it clearly.

Miss that step and you're not delegating. You're gambling.

So here's the question I asked myself in Spain, and the one I'd ask you: if your content ran for a fortnight without you, would you know, the first time you looked, whether it was truly yours or just more beige blandness?

That question is why this has taken 6 weeks to reach you. I could supervise the pipeline because I'd spent years learning the job it was doing. Most people I coach haven't had that runway. The gap between handing over the work and knowing whether it's any good is exactly where things go wrong.

So I've spent the time since I got back building an AI tutor that sits alongside my training programmes. It lets you ask questions of everything I've ever taught and get a straight answer back, the kind of answer I'd give you myself if you rang me up. It's going into beta with my existing students shortly, because I'd rather prove it with them than launch it blind.

If you'd like early access when it opens, join the waitlist. You'll be first in line to see what it answers, with a full walkthrough when it's ready.

Know the job before you hand it over. Hasta luego,

Paul


Want the whole stack? I ran 8 tools to make this work, and rather than break the story with each one, I've written the full breakdown, plus how I kept client data out of it, as a separate piece you can read in your own time.


Get Letters in your inbox

Keep reading