Content on 6 Platforms While You're Away: The Full Stack Breakdown


A technical report on the August 2026 atomisation pipeline, including where it failed
Paul Anderson, The Social Media Hotelier
Ad disclosure. The links to Flodesk, Eden and Descript in this paper are affiliate links. If you sign up through them, I may earn a commission, at no extra cost to you. Flodesk gives you 25% off your first year through my link, and Descript gives you 50% off the Creator plan for 2 months. I use all 3 tools in the pipeline described here, and I'd describe them the same way without the commission.
Here for the AI tutor rather than the engineering? I'm building an AI tutor that answers from everything I've taught. Join the AI tutor waitlist.
Abstract
In August 2026 I ran an AI-assisted content pipeline that published 1 piece of content a day to 6 platforms (Instagram, Facebook, LinkedIn, Threads, Substack Notes and X) for 16 days while I was on a family holiday in Spain, 5 to 19 August. The pipeline mined my own long-form archive, drafted and graded new content through a zero-sycophancy critic loop, rendered carousels by code, and scheduled everything through a single platform. It held. It also failed once, in a 48-hour window, in a place none of my checks were looking: the delivery layer. 4 images shipped at between a fifth and a third of their intended resolution, and my verification step reported every one of them as correct.
This report documents the architecture, the toolset, the quality controls, the failure and its root cause, and the results, including the ones that don't flatter the experiment. It is a single run by a single operator, and I've written down its limits as carefully as its findings.
Keywords: AI content pipelines, human-in-the-loop, verification, short-term rental marketing, 10-80-10 Protocol
How to read this
- If you host and want the plain version, read sections 1, 6 and 9.
- If you want to build something similar, read sections 2 to 5.
- If you only want the single most useful idea, read section 6. It explains how my automated check reported every post as fine when 4 images had gone out at between a fifth and a third of their intended size, because it only checked that each file had arrived intact, not how big the picture was.
1. Introduction
I coach short-term rental hosts on using social media to drive guests to booking pages. This report describes an experiment in automating that work. It wasn't a growth experiment. Its purpose was to test the automation of a theory I've been developing: the 10-80-10 Protocol.
The Protocol divides AI-assisted work into 3 parts:
- The first 10%, Clarity. The Commander's Intent. A human decides who the content is for, what keeps them up at 2am, what success looks like and what they want to hear.
- The middle 80%, the making. Drafting, formatting, creating. AI has commoditised this part.
- The last 10%, Human Refinement. The Command Inspection. A human reads back what came out and decides whether it's theirs or beige slop, then approves it.
Research question. If as much of the Protocol as was technically possible was automated, would the process hold for 16 days while I was overseas with no internet connection?
Scope. The test was designed to assess whether the process held. It wasn't designed to optimise content performance, and the results in section 7 should be read with that in mind.
2. System overview
The pipeline ran on a written standard operating procedure (SOP) that grew from v1 on 22 July to v4 on 23 August. By v4 it was 440KB and roughly 4,060 lines, navigated through an index rather than read end to end.
The pipeline ran in 9 stages, in this order:
- Sources. Blogs, emails, video and calls.
- Draft. Long-form, with the title written last.
- Grade. The Critic Loop: 9/10 or better on every dimension.
- Atomise. Self-contained short-form ideas.
- Platform text. For each atom, 4 text posts (X, Threads, Facebook and Substack Notes) and 2 captions (Instagram and LinkedIn).
- Render. Carousels built by code (Python).
- Automated gates. 3 automated checks.
- Schedule. Through Eden, driven in Chrome.
- Verify. Read back every post from Eden.
A gate is a point where nothing moves forward without my approval. There were 9:
| Gate | What I approved |
|---|---|
| 1 | Source material |
| 2 | Long-form draft |
| 3 | Atom concepts |
| 4 | The call to action for each atom |
| 5 | The text posts for each atom |
| 6 | Carousel slides |
| 7 | Carousel captions (6 and 7 reviewed together in one sitting) |
| 8 | Upload through the scheduling tool |
| 9 | The post-schedule report |
Two rules governed the gates. Concepts were approved before calls to action, and calls to action before drafting. Drafting was allowed to run ahead of a pending gate. Scheduling never was.
3. The corpus
The principle was simple: poor inputs lead to poor outputs. I wanted the AI producing content that looked, sounded, felt and smelt as close to my manually-created work as possible, because the closer it got, the smaller the last 10% would be.
| Source | Volume | How it was accessed |
|---|---|---|
| Coaching calls, webinars and course modules | 350+ hours; roughly 300 transcripts in a master index | Fathom and Descript, via MCP |
| Email newsletters | 332 lifetime sends, 830,000+ deliveries; top 60 ranked by unique opens | Flodesk, via MCP |
| Blog articles | 32 long-form pieces, September 2023 to July 2026 | Scraped and individually assessed |
| Digital training programmes | 50+ hours, plus the prep documents and script outlines behind them | Transcripts and documents |
| Podcast appearances | 8 transcripts | 1 pulled manually from YouTube using Glasp |
| Ideal client | 1 granularly defined persona | Defined by the author |
Two findings from the email data shaped everything downstream. Subscribers were most likely to open emails about platform news, but most likely to click through a link in the email when it told a story. And the best at-scale send, "Your polished Instagram is now a red flag.", reached a 63.0% open rate.
From this corpus the mining pass produced 102 candidate atoms against a minimum ask of 45 (3x cover at 1 per day). These were shuffled into an 87-atom working deck and worked strictly top-down.
4. The stack, in the order I reached for it
8 tools did the heavy lifting. A few small helpers (Glasp, a Gmail connector for reply data, and a real Chrome browser, covered below) sit around the edges.
4.1 Claude: the writing partner. It analysed the input, defined my voice, atomised the teachings and drafted the outlines. It did most of the heavy lifting.
4.2 Flodesk: the email archive. I set up an MCP, which is the software equivalent of a travel adaptor: it lets an AI talk to other software. It pulled all 332 emails and their performance stats so I could see which topics you open, which links you click and, via the Gmail connector, which ones got personal replies. If you'd like to try it, Flodesk is worth a look, and my link gets you 25% off your first year.
4.3 Descript: the video archive. I've used it for years to record and edit my training videos, so the transcripts were already there. A second MCP pulled them into the mix. If you'd like to try it, Descript gives you 50% off your first 2 months of the Creator plan through my link.
4.4 Fathom: the coaching calls. Fathom is an AI note-taker that joins your video calls and produces a recording, a transcript and a summary. For me that meant years of replay videos and time-stamped summaries. When I remembered it also produced transcripts, a third MCP made them available. Then Claude flagged something I hadn't considered: those calls are full of clients' names, their businesses and, in some cases, their finances. Section 9 covers what I did about that.
4.5 Fable, Opus and Sonnet: the models behind the extraction. Fable did most of this work, and I capped its budget at £50. It designed the extraction scheme, ran the sanitisation passes on the coaching-call transcripts, extracted my teaching from them and audited the results. I also used Opus (the most recent version available at the time) for part of the work, and Sonnet for some of the sanitisation. All of it ran inside a restricted folder (section 9).
4.6 Python: the graphics. I'm not a coder, and this was entirely Claude's suggestion. I described the graphics I wanted, it wrote the code, and I eventually gave it access to my Mac's terminal to run it. The rendering script used PIL to produce every 1080x1350 slide from locked constants:
- navy background, cream text, Raleway only
- 64pt headline and 64pt subline on every slide
- fixed margins, a radial overlay, a brand footer and a swipe arrow
- fit problems fixed by rewording, never by resizing type
- slide 2 always built as a second cold entry point, because Instagram serves slide 2 on a post's second pass through the feed
86 slide files were rendered in 4 days. One slide went through 13 versions, and a typical carousel took 4 to 8 render rounds. One atom took 8 proof PDFs in a single evening. No pixel was touched by me.
4.7 Unsplash: the backgrounds. 107 images in a brand assets folder.
4.8 Eden: the command centre. The critic loops, the analytics, the approval gates and the scheduling. This is where I stayed in the loop without being in the way. If you want to build something similar yourself, you can start with Eden.
A note on scheduling. The direct upload APIs failed in the sandbox: blocked storage endpoints, and base64 relays that weren't reliable above roughly 16,000 characters. The route that worked was driving a real Chrome browser through Eden's scheduling interface: new draft, upload PNGs in slide order, carousel format, per-platform customised captions, natural-language scheduling, and a re-check of the platform toggles every time, because the toggles carry over between drafts. Cadence settled at 1 atom a day, every platform in the same slot, 18:08 UK time.
None of it turned out to be exotic. That was the surprise.
5. Quality control
5.1 The Critic Loop
Every long-form draft went to a content critic in Eden with a strict rubric. It graded every dimension (hook, voice, specificity, payoff, subject line, sustained read, clear takeaway), ran an adversarial pass before writing any score, tied every criticism to a specific line, revised, then re-graded cold.
The exit rule was 9/10 or better on every dimension. It was a floor on each dimension, not an average, so one weak dimension failed the draft however strong the rest was.
The first grade on the first newsletter came back at 96/100. It felt inflated, so I told the critic to drop the flattery and grade again. The same draft scored 72. We ran the loop until it shipped, 5 versions later, at an honest 90.3. Zero sycophancy became a permanent written rule across the whole chain.
A second finding followed. 2 hooks scored 9/10 on the general rubric, then 39/100 and 37/100 when graded cold on 10 hook-specific dimensions. After that, every hook surface was graded separately: slide 1, slide 2, the openers of the 4 text posts, both captions, the title, the subject line and the preheader.
5.2 The 3 automated gates
Before I saw anything, the work passed through 3 automated checks. Each was a short program that tests the work against a fixed rule, with no human judgement involved:
- A repeated-phrase checker. It compares every piece of an atom and flags any run of 4 consecutive words that appears on more than one platform (a 4-gram is simply a run of 4 words), so no 2 platforms shipped the same sentence.
- An orphan-word check. It runs each slide's text through the renderer's own line-wrapping at 64pt and flags any line left with a single word stranded on it.
- A banned-word and AI-tell sweep: em dashes, hashtags, "quiet", "-ize" spellings.
These checks mattered because human attention wasn't reliable. A banned word crept back in after being banned 3 times, and only the automated sweep caught it. In the words of my own working notes: "Human vigilance demonstrably does not hold across eight atoms."
5.3 One version per platform
The rule that made the short-form feel native: each atom's text posts were 4 genuinely different versions, one each for X, Threads, Facebook and Substack Notes, written with 4 different hook types, not one text trimmed 4 ways.
| Platform | Constraint that shaped the copy |
|---|---|
| X | 280 characters a segment; the whole first tweet is the hook (the data showed a 95% view drop from tweet 1 to tweet 2) |
| Threads | Hard 500-character server cap, learned from a live rejection at 742 characters; the most conversational register |
| Story register; link in the first comment, never the caption, because Facebook throttles caption links | |
| Substack | Register close to Threads; "subscribe" not "follow"; 6-image cap, so 8-slide carousels go text-only |
| Roughly 125 characters before "more"; warm; link in bio | |
| Roughly 140 characters show before LinkedIn cuts the post off behind "see more"; outcome named faster; jargon banned |
The universal rule: design to "the first two lines", not to a character count. Captions never repeat slide text, and a caption's hook type must differ from slide 1's.
6. Failure analysis: the delivery layer
6.1 Timeline
- 8 August. 1 image went out with a grey block across its lower two-thirds: a genuinely corrupted file, live on 3 platforms for hours.
- 9 August. 4 posts shipped at 194x243, 237x297, 324x405 and 216x270 pixels against 1080x1350 sources. 3 of them carried the same text, posted repeatedly over 3.5 hours, because failed attempts were retried without the earlier attempts ever being cancelled.
My verification stage reported all of it as "byte-verified" and correct.
6.2 Root cause
The forensic diagnosis, written on 31 August, found that the direct upload path silently truncates payloads above roughly 2,900 bytes. Something in the hand-off compressed each image iteratively until it fit under that ceiling. The evidence was 4 different pixel sizes, every byte count just under 2,900: the signature of a compressor hitting a byte budget.
6.3 Why verification passed
The "byte-verified" check compared bytes sent with bytes received. Both numbers were correct, and they matched, so the check passed on ruined files. It was measuring the wrong property.
A check that measures the wrong property will pass on a ruined file. Here, the bytes sent matched the bytes received, and neither number said anything about whether the image was usable.
The AI couldn't spot the difference. It had no eyes.
6.4 Corrective action
- The check now asserts pixel dimensions, with a size floor.
- Metadata that is missing counts as a failure, not a skip.
- Still to do before the next run: make the post-schedule report assert 1 post per platform per day, which targets the missed platforms in section 6.6.
6.5 Recovery
On the morning of 10 August a Manual Posting Guide was generated for the remaining 12 days: captions pulled verbatim from Eden's own records, pointing at the full-resolution source files on disk, with explicit instructions not to reuse anything already attached in Eden, because every one of those images was degraded.
I deleted the broken drafts and re-posted by hand from the Eden mobile app. The app refuses to schedule Facebook, Instagram and LinkedIn together with Threads in one draft, so every day took 2 sessions, and X's 4-image cap meant larger carousels had to be trimmed for that platform.
Cost of the fix: about 2 hours 15 minutes to 2.5 hours, spread across 2 days.
6.6 Audit
An audit of all 44 assets across 9 to 21 August found the damage confined to 9 August plus the 8 August glitch: 40 of 44 clean at full resolution. The remaining gaps were less visible: 10 August reached only 3 of 6 platforms, Threads and X were missed on 2 days, and nothing was posted to X on 6 August.
Observation. This was a systems failure, not a failure of the principle. Both 10s held: the Clarity was mine and the Refinement was mine. The break happened in the delivery layer, which sits downstream of the 10-80-10 itself and after I'd signed off, so it was a technical fault rather than a flaw in the Protocol.
7. Results
Window: 4 to 19 August, 16 days. Sources: Eden analytics (post level), Meta Business Suite, Threads Insights, LinkedIn native analytics, the public X profile and Substack stats. Captured 31 August to 1 September 2026. Percentage changes are against the previous 16 days.
| Platform | Headline figures |
|---|---|
| 3,746 views (up 19.4%); 596 accounts reached (up 105.5%); 88 profile visits (up 69.2%); 1 link click; 8 follows against 10 unfollows (net -2) | |
| Gained 25 followers; impressions fell 50% to 518; 138 members reached | |
| 294 views (down 82.9%); 5 to 52 views per post; 0 follows | |
| Threads | 348 views; followers 548 to 549 |
| Substack Notes | 22 likes across 16 notes; followers 112 to 113 |
| X | 110 views across every post; 0 likes |
Observations
- Instagram carried the reach, but the doctrine's own metrics show attention without action. 8 follows from 88 profile visits is a 9% conversion rate. My doctrine sets a floor of 20%.
- The platforms wanted opposite things. On Instagram, rule-and-list posts took the top 4 slots and story posts did worst. On LinkedIn it flipped: the RAF Selection story was the best post of the run, at twice its baseline.
- The 2 best Instagram posts came on 4 and 5 August. Nothing after 8 August reached 200 views.
- A hand-posted Instagram Story outperformed the scheduled feed. "Increíble... otra vez.", posted on 13 August, took 349 views and 271 reach, more than any scheduled feed post reached in the fortnight. Across the 17 Instagram Stories I posted by hand, the average was 56 views.
8. Other failure modes
The image corruption was the headline failure, but not the only one. Each of these was logged.
| What happened | What came of it |
|---|---|
| The AI invented things: a workshop that didn't exist, a coinage attributed to me, doctrine written in my voice that I never said. A single prompt audit found 21 unsourced items. | Every claim is provenance-tagged and prompts are audited line by line. The failure was named in writing as the most damaging on the project. |
| The AI scheduled atom 2 before atom 1, before I'd reviewed it. The SOP now calls this the worst process failure of the rollout. | Scheduling now never runs ahead of a pending gate, and atoms publish in strict numerical order. |
| After its context was compacted, the AI reconstructed an approved caption from memory, slightly wrong. I caught it live. | New standing rule: "do not reconstruct approved copy from memory if there's any doubt." |
| My app crashed every time I saved the 394KB master doctrine file. | The file was split in two, loaded as two separate documents, then merged. |
| An experiment to auto-cut video clips with Descript's AI burned credits faster than it returned value. | Shelved on 4 August. |
One finding I keep for perspective: the newsletter about AI making everyone's marketing sound the same was itself flagged by an AI detector as "100% definitely AI writing". I've logged it as a gap, and filed it under irony.
9. Data governance: keeping client data out
The coaching call archive was the richest source in the corpus and the riskiest. Those calls contain clients' names, their businesses and, in some cases, financial details. Claude flagged it when I connected the archive.
The scheme that came out of that was built to keep everything contained. It has 9 parts.
9.1 A closed environment
The work ran inside a closed environment. I used Claude's Dispatch tool to plan the approach, and it recommended setting up a dedicated agent for the job, so I asked it to create one running Fable. Most passes used Fable, with some on Opus and some on Sonnet. The transcripts sat in a restricted folder on an external hard drive, and that folder was the agent's sandbox. Neither the raw nor the sanitised transcripts were ever shared publicly or outside my own accounts. Real identities were kept out of the AI's reach throughout. It worked with codes, never with names, and it couldn't turn one back into the other.
9.2 Pass 1: replacing names with codes
The first pass ran across 303 transcripts. All known client and business names (256 codes in total) were replaced with typed codes, including many spelling and casing variants. Emails and email-like spoken mentions were redacted, URLs removed, and currency figures rounded into ranges. Every file was renamed to an anonymous sequential ID, and speaker turns were restructured into a uniform format.
9.3 Typed tokens, not deletion
Names were replaced rather than deleted. A person gets a code. A business gets a token linked to its owner's code, which preserves who owns what while killing identifiability. Towns and landmarks become region tokens, URLs become tokens, and verbatim-quoted blog or post titles are paraphrased because they're googleable. Where a combination was realistically searchable, such as a specific race name beside a finish time, both were generalised.
9.4 Repeated passes from different angles
One pass was never treated as enough. Free-flowing speech has misspellings, concatenated words and one-off nicknames that pattern matching can't catch with certainty. So I ran the agent over the transcripts repeatedly, each time asking it to approach them from a different angle: business names, money figures, regulatory details, then individual names. Rounds like these found residual identifying details that earlier passes had missed: a forename rendered 2 ways, a surname fused to the word beside it, a business name run together with the word before it, one wholly phonetic mis-rendering. After several fix passes, a search across 147 known full names found zero residual hits. That isn't a guarantee of zero edge cases, and I've never claimed it was.
The errors ran in both directions. Transcription garbles were mistaken for names (a design comment about a font was flagged as a business), and idioms and public figures were flagged as clients. I kept a list of confirmed false alarms so they stopped recurring.
The extraction pass then found 21 further residual identifying details beyond those already fixed, which is why repeated passes were worth the effort.
9.5 The refusal to guess
Where identity was ambiguous, the pass stopped. A common first name that clearly belonged to several different real people across about 24 files stayed unmapped, because a blanket merge would have folded different clients' private histories together. Two live group calls were left untouched for the same reason. The standing discipline is "flag rather than force": an honest unresolved flag has repeatedly beaten a confident guess.
9.6 Pass 2: sanitise the output, not just the input
This is what makes it more than find-and-replace. An AI's own paraphrases can leak identity, and a fully anonymised but distinctive anecdote is still recognisable to the person who lived it and to their peers. Nothing moves into the product layers until the output pass is marked verified. Anything flagged as needing a composite can only appear customer-facing as part of a composite of 2 or more clients, or as a synthetic example, never from a single source.
A locked layer rule sits beside it: product material may only be built from the distilled layers, never directly from the anonymised excerpts. Fable's stress-test of my plan flagged this as the single biggest way the whole pipeline gets undone.
9.7 Consent is a separate gate
Redaction answers whether someone can be identified. Consent answers whether they agreed to appear. Consent is given person by person, and only for what that person agreed to. The default is that transcripts can inform what I teach, but they can't be reproduced, attributed or cut into short-form video.
9.8 What redaction can't solve
Some material is sensitive whatever anyone agreed to. Some stories identify a person even with the name taken out, because the combination of details belongs to them alone. The fix I settled on is a test: strip the identifying composite and keep what makes the point work. "A Lake Como villa at €25,000 a week" identifies nobody, because Lake Como has many such properties, and that plurality is what makes the teaching point work. If what's left still locates the person, it stays members-only.
A related risk is synthesis. One flagship case briefly carried a specific result figure that felt natural to write but never appeared in the source. The ledger had deliberately redacted the real number. Writing a figure because it reads more convincingly than a vague one can reintroduce exactly the specificity a source stripped. The fingerprint is also cumulative across assets: 3 assets that each keep a different harmless detail of the same story can re-assemble the identity between them.
9.9 The rule I hold
None of this material enters a product or an AI knowledge source until it's been sanitised and I've checked a sample myself.
10. Limitations and threats to validity
- n = 1. One run, one operator, 16 days, one niche.
- Confounded variables. The 2 best posts came before the incident, and they were also the strongest concepts. Format and fault can't be separated from this data.
- Missing data. X follower change needs Premium. Facebook Story insights were switched off at the time and aren't recoverable. Instagram Stories from 4 August and 16 to 19 August had expired from Meta's 28-day window. Per-post profile visits and link clicks are phone-only.
- Excluded data. The 9 August duplicate reposts never resolved metrics in Eden and are excluded from every count.
- Mixed measures. LinkedIn's 25 followers is a gross gain (LinkedIn doesn't surface unfollows), while Instagram's -2 is net. They aren't directly comparable. LinkedIn's native 518 impressions are account-level and unique; Eden's 373 are per-post sums.
- A modest absolute scale. Most of these numbers are small. They show the shape of a result, not statistical proof.
11. Discussion
What I'd keep. The SOP survived its author's absence. It posted daily to 6 platforms while I was in another country, failed in exactly 1 window, and every failure produced a written rule that makes the next run stronger. The 10-80-10 split held: my thinking at the front, my judgement at the back, and the 80% in the middle done at a volume no individual could sustain by hand.
What I'm changing for the next run. These are my decisions, taken from the analysis.
- Prioritise profile visits, follows, link clicks and profile conversion, while still tracking views.
- Give Facebook a native post or nothing.
- Split atoms by platform, not just captions: instruction-led for Instagram, narrative-led for LinkedIn.
- Front-load the strongest concepts and re-run them.
- Explicitly permit ad hoc posting whilst the strategic, automated content is published.
- Make the post-schedule report assert 1 post per platform per day.
The wider point. The skill that held this together wasn't prompting. It was knowing enough about the job to supervise the work I'd handed over. When the machine couldn't see the damage, I could. That's what the Protocol is for, and it's the gap I'm trying to close for the hosts I coach: the distance between handing over the work and knowing whether it's any good.
12. Where to go next
The pipeline above is mine, built on years of knowing the job it was doing. Most hosts I coach haven't had that runway, so I've spent the time since I got back building an AI tutor that sits alongside my training programmes. You ask it a question and get a straight answer from everything I've taught, the kind of answer I'd give you myself if you rang me up.
In the meantime, you can read my free blog articles as I move them into Eden, in a public archive called Letters.
Read the free blogs in Letters
Appendix A: The repeatable shape, in 10 steps
- Mine the long-form assets you already own.
- Draft one piece at a time, with the title written last.
- Grade it through a critic with a strict rubric and a zero-sycophancy rule.
- Atomise it into self-contained short-form ideas.
- Trim in person, at named approval gates.
- Write each platform its own version.
- Render the carousels by code, from locked constants.
- Run automated quality gates before you look at anything.
- Schedule through one platform.
- Verify by reading every post back out, and show the evidence, never a bare claim of success.
Appendix B: Glossary
- Atom. A self-contained short-form idea, tagged by format (carousel, quote card, story, template, stat, testimonial).
- Critic Loop. The grade-revise-regrade cycle described in section 5.1.
- Gate. A point where nothing moves forward without my approval.
- MCP. The software equivalent of a travel adaptor: it lets an AI talk to other software.
- SOP. The written standard operating procedure the pipeline runs on.
Appendix C: Source records
This report draws on: the SOP versions 1 to 4 and its index; the process log; the atom shortlist for the 8 to 22 August buffer; the atom pipeline session checkpoint; the content sourcing research; the Manual Posting Guide for 10 to 21 August; the image resolution loss diagnosis of 31 August 2026; the Eden inventory of 31 August 2026; the performance analysis for 4 to 19 August; and the August Claude conversation exports.
Prefer to read? Browse the free blogs in Letters.
