--- title: "I opened the client's 'Drafts' folder. There were 47 posts in it" url: "https://www.mikefluff.com/en/blog/content-pipeline-why-it-stalls" locale: "en" author: "Mike Fluff" updated: "2025-10-08" description: "AI wrote fifteen posts a week and the school published one. Where an AI content factory stalls, and the five changes that got it publishing again." tags: - "content" - "ai" - "workflow" alternates: - locale: "en" url: "https://www.mikefluff.com/en/blog/content-pipeline-why-it-stalls" markdown: "https://www.mikefluff.com/en/blog/content-pipeline-why-it-stalls.md" - locale: "ru" url: "https://www.mikefluff.com/ru/blog/content-pipeline-why-it-stalls" markdown: "https://www.mikefluff.com/ru/blog/content-pipeline-why-it-stalls.md" - locale: "pt" url: "https://www.mikefluff.com/pt/blog/content-pipeline-why-it-stalls" markdown: "https://www.mikefluff.com/pt/blog/content-pipeline-why-it-stalls.md" - locale: "es" url: "https://www.mikefluff.com/es/blog/content-pipeline-why-it-stalls" markdown: "https://www.mikefluff.com/es/blog/content-pipeline-why-it-stalls.md" --- # I opened the client's 'Drafts' folder. There were 47 posts in it > **In short:** It stalls because only the drafting is automated and the other ten or so steps stay manual. At an online school AI wrote fifteen posts a week and one got published. We paused generation, deleted 31 of 47 drafts, put every stage on one board, made the prompt a shared artefact and capped manual work at twenty minutes a post. Two months later they published four a week. There was a call. The client — an online school, 40 people on the team, living off regular Telegram and Instagram content. The CMO called and said: "our content factory broke. Help me fix it." I sat down to look. An hour in I opened their Notion folder called "Drafts." There were 47 posts in it. Of those, older than three weeks — 31. Older than two months — 14. One, the oldest, was marked "ready to publish" — **a year and a half ago**. That post was about a trend which, by the time I was looking, was no longer a trend. Presumably it had been a trend when written. Maybe. Nobody remembered. — So, — says the CMO. — We do publish, though. About one post a week. — Does your SMM person write one a week? — No. AI writes about fifteen a week. Right. So fifteen a week were being written. One a week — published. The other fourteen — settled into "Drafts," where they quietly died of old age. That was their "content factory." The AI worked. The factory didn't. --- I've seen this picture at eight or nine clients since. Different scale (one a week vs five a day), but the same structure: **generation is automated, everything else isn't**. The result is a **bottleneck in the manual part**, and as soon as that shows up, the whole pipeline stops. Drafts pile up, pain piles up, and in a month or two the team stuffs the thing in a corner so they don't have to look. The typical response: "well, the AI must write badly." Wrong. The AI writes well enough. The problem is that **nobody called "content factory" all the other stuff** needed for normal operation. And the other stuff is roughly 80% of the work. --- A factory isn't a machine. It's a **workshop**. A workshop has ten machines, a raw-material warehouse, an assembly floor, QC, transport, and a person who mops the floor once a week. If you only have a machine, that's a **stall**, not a factory. A stall scales linearly in effort, not by two orders. Here's what usually is and isn't automated in an "AI content factory." **Automated:** draft generation. Maybe headline generation. Sometimes — visual generation. That's it. **Not automated:** Idea sourcing. Topic research. Editorial judgement ("is this actually what we want to say?"). Fact-checking. Adaptation to platform format. Approvals — with legal, for instance, if the product is in a regulated space. Publication scheduling. Actual publication. Response collection. Analysis of what worked. Feeding analysis back into the next cycle. Add it up — **ten steps**. Automated — **one or two**. And I haven't even added the minor ones: "write the post in three tones for three channels," "rework to match our visual standards," "comply with our internal terminology rules." Three more steps. --- Three things I see at every client like this. **First — "I have lots of ideas, I'll process them later."** No. You won't. Ideas **go stale**. They're tied to the moment — a trend, a news cycle, a season, an audience mood. An idea that sat for three weeks is **no longer your idea**. It's something that resembles it. The idea funnel must be **wired to a publishing schedule**. Not "we collect ideas, then decide." Instead: idea → topic → draft → publication, **with dates locked to each other**. An idea without a date is a museum. **Second — "AI writes, we'll fix it up."** This works exactly until "fix it up" starts to equal "rewrite." That moment arrives faster than you think, and it arrives because you framed the prompt as "write a post about X in our brand voice" without specifying what "our brand voice" is and how it differs from "the voice of any similar brand." I have a hard rule: **if editing takes more than 30% of the time it would take to write from scratch, the prompt is broken**. You fix it by changing the prompt, not by adding "another editorial pass." Run a couple of posts through the [AI text checker](https://www.mikefluff.com/ai-text-checker). If every one of them carries the same stock phrases, that is another sign of a broken prompt, and the prompt is again the thing to fix. > **A simple test** > > Give the editor ten posts from your pipeline. Have them time editing. If total > — two hours, and writing from scratch would have been six — the factory works. > If 1.5 hours vs 2 from scratch — you have **negative economics**. The factory > is **costing you**. **Third — "we use five tools, not one system."** Notion for ideas. Google Docs for drafts. Telegram for approvals. Linear for tickets. A spreadsheet for analytics. Each tool — good, on its own. Together it's **five extra clicks** per unit of content. Five times twenty publications a month — a hundred clicks. A hundred clicks is an hour of your attention **every day**. In a content factory a key metric is **friction per unit**. If it's high, the team gets tired. A tired team loses quality. Lost quality means content stops doing its job. Back to the "we publish but we get no leads" place. --- What I did with the client and their 47 drafts. First — **shut down the AI generation**. Fully. Two weeks. Because **that's the source of the blockage**. You can only clear a backlog by stopping adding to it. Second — **cleared drafts**. Not "edited 47." **Deleted 31.** Because 31 had lost relevance. Left 16. Of 16 — 10 published inside two weeks. 6 — kept as a buffer, but with concrete dates. Third — **rewrote the process**. From idea to publication, **one board** (they ended up on Airtable, but anything works). One record = one publication = all stages on one row. Idea, author, draft, editor, approval, date, channel, results. No "check Notion, then Docs, then Telegram, then the spreadsheet." One screen. Fourth — **the prompt became part of the system**. Previously each writer called ChatGPT with "write a post about X in our brand voice" their own way. Now the prompt is an artefact, living in Airtable next to templates. Update it — everyone writes in the new way. A small-seeming detail, but it's what gives the team **consistency**, without which a content factory can't be a factory. Fifth — **manual-work budget per unit**. No more than twenty minutes per post from draft to publication. More than that, investigate why. **Budget includes editing.** If editing eats fifteen minutes — that's five for everything else, time to change the pipeline. --- Two months later, instead of "one post a week," this client was publishing four a week. Not fifteen, the way the old factory "produced." Four. But four **published** and **doing their job**, not fifteen rotting in Notion. That's the key metric no one tracks: **ratio of published to generated**. Before the intervention theirs was 1 to 15. Two months in — 4 to 5. Not scale. **Recovery.** --- A content factory isn't "AI that generates posts." It's a **workshop** where AI is one machine. If you bought the machine and didn't build the rest, you'll end up where my client did: 47 drafts in a folder and a CMO who doesn't understand why volume goes up but reach doesn't. Machines are cheap. A workshop is expensive. But only the workshop produces. A machine on its own produces **parts**, which nobody then assembles. **[Content factory course](https://www.mikefluff.com/en/content-factory-course)** — Eight modules, $19. On how to build the workshop, not just the machine, and how not to let it break in week three.