Three weeks ago, I set out to build something ambitious: a fully AI-assisted content pipeline that could research trending topics, write long-form articles, generate cover images, format content for six social media platforms, and submit everything as drafts — automatically.
The vision was simple. Wake up, tell the AI what to build, review the output, hit publish. No blank page anxiety, no hours spent reformatting the same content for Twitter, Instagram, Facebook, LinkedIn, TikTok, and Medium. Just a daily briefing, a review, and done.
Here's what actually happened.
The Stack
The pipeline runs on a surprisingly lean set of tools:
Research: Web search to find the latest AI news, model releases, and industry trends. The AI identifies the highest-value topic for the day, cross-references multiple sources, and builds a factual foundation before writing.
Writing: A large language model handles the actual drafting — pillar articles for the website (1500+ words), then repurposed versions for each platform. Each platform gets its own format: Twitter threads, Instagram carousels, Facebook stories, TikTok scripts, Medium articles, LinkedIn technical posts.
Image generation: Cover images are AI-generated based on the article theme, then uploaded to cloud storage (Supabase) and linked to the article automatically.
Submission: An Automation API receives the article payload — title, slug, content, cover image URL — and creates it as a draft in the CMS. The published: false flag ensures nothing goes live without human review.
Workflow automation: The entire sequence is codified in workflow files that the AI follows step-by-step. Think of them as playbooks: "Step 1: research. Step 2: outline (stop for approval). Step 3: write. Step 4: generate images. Step 5: submit drafts."
The total time from "go" to "drafts ready for review" is about 15-20 minutes, depending on article complexity. For comparison, manually researching, writing, formatting for six platforms, and uploading images used to take 4-6 hours.
What Works Really Well
Research is genuinely good. The AI consistently finds relevant, timely topics and synthesizes multiple sources into coherent summaries. It catches trends I would have missed. When five Chinese AI labs launched models in the same two-week window, the AI connected the dots across announcements that I might have processed individually.
First drafts are 70-80% there. The pillar articles come out structured, well-researched, and surprisingly readable. They're not perfect — they tend toward a particular style that needs human editing for personality — but they eliminate the hardest part of writing: starting from zero.
Social repurposing is the biggest time saver. Taking a 1500-word article and manually reformatting it for Twitter (280-char limit per tweet), Instagram (carousel slides), TikTok (30-second script), Facebook (conversational tone), Medium (SEO-optimized), and LinkedIn (technical audience) used to take longer than writing the original article. The AI handles this in seconds and does a surprisingly good job of adapting tone for each platform.
The human-in-the-loop principle prevents disasters. Every piece of content is created as a draft. I review everything before publishing. This has caught inaccuracies, tone issues, and formatting problems that would have been embarrassing at best, harmful at worst.
What Doesn't Work (Yet)
Platform formatting is a nightmare. This was my biggest lesson, and it cost me several frustrating hours. Medium doesn't render Markdown. When the AI writes ### Section Title and **bold text**, Medium displays those as literal characters: ### Section Title and **bold text**. I had to completely rebuild the Medium section of the workflow to output clean plain text with a separate formatting guide.
The same issue applies differently to every platform. Twitter ignores formatting entirely. LinkedIn renders line breaks inconsistently. Instagram strips everything but emojis and hashtags. Facebook has its own rich text quirks.
The lesson: each platform is a unique rendering engine with unique rules. You can't write once and paste everywhere. The AI now has a formatting rules table that specifies exactly what syntax is allowed on each platform. It took three iterations to get right.
Tone consistency is hard to maintain. The AI writes well, but it has tendencies. It over-uses certain phrases ("Here's the thing," "Let me be direct"), tends toward a slightly formal register, and sometimes sounds more like a press release than a person. I've been iterating on the writing tone instructions, and it's getting better, but achieving a genuinely human voice requires editing that no prompt engineering has fully solved.
Fact-checking cannot be automated. The AI gets facts right most of the time, but "most of the time" isn't good enough for published content. I've caught incorrect benchmark numbers, misattributed quotes, and dates that were off by a day or two. The human review step isn't optional — it's essential. Anyone building an AI content pipeline without human fact-checking is publishing a liability.
Image generation is hit-or-miss. Cover images are sometimes stunning, sometimes generic. The AI doesn't have a persistent understanding of brand aesthetics — each generation is independent. I'm exploring ways to enforce visual consistency, but for now, I sometimes regenerate images 2-3 times to get something that matches the article's energy.
Five Concrete Lessons
After three weeks of daily production, here's what I'd tell anyone building a similar system:
1. Codify everything into workflows. The magic isn't in the AI's capabilities — it's in the structured playbooks that tell it exactly what to do, in what order, with what constraints. Workflow files are the real product. The AI is the execution engine.
2. Build for "draft only" from day one. Never give an AI pipeline the ability to publish directly. Always insert a human review step. The quality gap between "impressive first draft" and "ready to publish" is where your credibility lives.
3. Platform formatting rules save hours. Document exactly what rendering each platform supports. Put it in the workflow. Make the AI follow it. This prevents the single most common and annoying type of error.
4. Iterate on tone constantly. Your writing voice is the hardest thing to capture in a prompt. After every review, note what you changed and feed it back into the system. The tone instructions in my workflow have been rewritten five times.
5. Time savings are real, but front-loaded. Setting up the pipeline took several days of intensive work — building workflows, debugging formatting, configuring APIs, training the AI on preferences. But once it's running, daily content creation genuinely takes 15-20 minutes of review instead of 4-6 hours of production.
What's Next
The pipeline isn't finished. Here's what I'm building next:
Scheduling. Right now I manually publish approved drafts. I want the system to queue approved content for optimal posting times on each platform.
Analytics feedback. Which posts perform well? Feed that data back into the AI to influence future topic selection and formatting choices.
Multi-language. I've started adding Vietnamese versions for Twitter and Facebook. The goal is full bilingual content for every platform, with native-quality writing in both languages — not translation, but separate articles written for each language's audience.
Visual consistency. A design system for AI-generated images — mood boards, color palettes, typography rules — so every cover image feels like part of the same brand, not a random generation.
The Honest Take
Building in public means being honest about the messy parts. The AI content pipeline isn't magic. It's a force multiplier — it makes a human writer 5-10x more productive, but it doesn't replace the human. Every article still needs editing. Every fact still needs checking. Every platform still has its formatting quirks.
But for a solo creator who wants to maintain presence across seven platforms with daily content? This system has changed the math entirely. The bottleneck has shifted from "I don't have time to create content" to "I need to get better at reviewing and refining content."
And honestly? That's a much better problem to have.
Follow along as I continue building and iterating on this system. The workflows, lessons, and code are all documented in real-time.