We Built an AI Content Pipeline That Runs 24/7 — Here's the Architecture

Every business knows content is the growth engine. Blogs drive SEO. Social posts build audience. Email nurtures leads. Video converts. The problem is not knowing what to do — it is doing it every day, at scale, without burning out a team.

We run a technology company in the chess industry. Two live web applications, six social channels, two blogs, outreach pipelines, and a directory that indexes over 500 entities. A content operation at this scale normally requires a content writer, a social media manager, a designer, and someone coordinating all of it.

Instead, we built an AI content pipeline that runs autonomously. Every day — weekends included — the system discovers topics, drafts blog posts, generates hero images, queues social content across every channel, and stages everything for human review. The pipeline has been running in production for months.

This is not a demo. It is not a proof of concept. It is the actual system that powers our daily content output. And it is the same architecture we build for clients at Early to AI.

Here is exactly how it works.

The content pipeline has five stages. Each stage is autonomous — it runs on a schedule without anyone triggering it — but every stage has a human review checkpoint before anything goes live.

Stage 1: Topic Discovery. Every morning, AI research agents scan the web for trending topics in our industry. They check news sites, official announcements, social media, and competitor content. The system picks three topics per day based on relevance, search volume, and freshness. Topics are logged to avoid repetition — the system checks what it has already covered and steers toward gaps.

Stage 2: Blog Drafting. For each topic, the system researches across five to ten sources, cross-references facts, and drafts a full blog post. Not a 200-word summary — a real article with structured sections, SEO metadata, internal links, and a clear editorial angle. The draft lands in a content queue for review.

Stage 3: Visual Asset Generation. Every blog post gets a hero image generated alongside the draft. The system uses AI image generation tuned to our brand guidelines — consistent color palette, typography, and visual style. For social distribution, the same pipeline generates platform-specific formats: square for Instagram, landscape for Twitter, vertical for Stories.

Stage 4: Multi-Channel Distribution. A single blog post does not stay on the blog. The system extracts key points, reformats them for each channel, and queues social posts across Twitter, Instagram, YouTube, TikTok, and email. One content event becomes six or more distribution events. Social posts are staged as drafts in our publishing system — not posted automatically.

Stage 5: Human Review and Publish. Everything lands in a review dashboard. Blog drafts, social posts, hero images, outreach emails — all in one place. The human reviews, edits if needed, and approves. The system publishes. The bottleneck is always approval, never execution.

The pipeline is not one monolithic application. It is a set of composable automation units — we call them skills — wired together with scheduled tasks, serverless functions, and live API integrations.

Claude Code Skills. The execution engine. Each skill is a self-contained automation unit: one handles topic research and blog drafting, another discovers new entities for the directory, another generates social content, another builds complete website profiles for potential customers. We have over twenty skills running in production. They execute interactively when we trigger them or autonomously on a schedule.

Firebase and Cloud Functions. The backend for everything. Firestore holds the content queue, the entity index, the CRM, the outreach pipeline, and the blog drafts. Cloud Functions handle server-side automation that needs to run regardless of whether anyone is at a computer — tweet fetching, content scoring, webhook processing.

n8n Workflow Automation. Running on a VPS, n8n handles email sequences, SMTP delivery, webhook routing, and scheduled jobs that need to fire reliably. Welcome sequences, drip campaigns, transactional emails — all automated through n8n workflows with HTML templates that match our brand.

Post-Bridge. The multi-platform publishing layer. Every social post — Twitter, Instagram, YouTube, TikTok — routes through Post-Bridge as a draft. We review in one interface and publish to all channels simultaneously. No logging into four different platforms.

Canva MCP. Static visual assets — Instagram posts, YouTube thumbnails, story graphics, promotional flyers — generate through the Canva integration. The AI creates designs using brand templates, exports them, and attaches them to the corresponding social drafts.

Playwright. Headless browser automation for site audits, deployment verification, SEO checks, and scraping JavaScript-rendered pages that traditional web fetching cannot handle.

Obsidian Vault. The knowledge graph. Strategy documents, CRM contacts, daily planning, operational memory — all in markdown files that AI agents read for context before executing tasks. The vault is the business brain. Firestore is the action layer. They never overlap.

The system runs on a combination of scheduled desktop tasks and server-side functions. Here is what a typical week looks like:

Daily (every morning): - Content research agents discover three topics, draft blog posts, and generate hero images - Entity discovery agents find new apps, tools, creators, and platforms — each gets a landing page, a companion blog post, and social distribution queued - Outreach agents scan for professionals who need websites, score them as leads, and for high scorers, build complete website profiles before any email is sent

Tuesday, Thursday, Saturday: - A Cloud Function fires server-side and fetches the top trending social content in our niche. It scores and filters automatically. No laptop required — this runs on Google Cloud regardless of whether anyone is working.

Tuesday and Wednesday: - Tweet fetch and quote fetch tasks populate the carousel content queue — source material for branded Instagram carousels

Sunday and Wednesday: - Rendering tasks produce finished Instagram carousels from the queued content and create Post-Bridge drafts for both brand accounts

Monday and Thursday: - Product marketing tasks generate social content for our SaaS product — tweets and Instagram drafts showcasing features

Tuesday: - A dedicated blog task publishes an SEO-optimized article on our SaaS product blog

The key detail: none of this requires someone to be at a computer. Desktop scheduled tasks fire when the machine is on and catch up on missed runs. Server-side functions fire regardless. The system assumes the human is busy and runs without them.

The real power of this pipeline is not any individual piece. It is the compound effect — every content event feeds multiple channels, and every channel feeds the next content event.

A single entity discovered by the directory pipeline generates: one SEO-rich landing page, one companion blog post, one tweet, one Instagram post, and one outreach email to the entity itself. That is five pieces of content from one discovery event.

A single blog post generates: the blog itself for SEO, a Twitter thread summarizing the key points, an Instagram carousel with the highlights, a potential YouTube script, and email newsletter content. One writing event becomes five distribution events.

Over time, this compounds. Our directory indexes over 500 entities. Each entity page builds domain authority and long-tail search traffic. Each blog post strengthens internal linking and topical authority. Each social post grows the audience that sees the next social post. Each outreach email leads to a relationship that generates future content opportunities.

The daily volume is modest — a few blog drafts, a handful of social posts, some outreach emails. But modest daily output times 365 days per year is not modest at all. Consistency at scale is the strategy. The system never takes a day off, never gets writer's block, and never forgets to post.

We are transparent about what the system actually produces because we think specificity builds more trust than vague claims.

500+ entities indexed in our industry directory. Each with a dedicated landing page, structured data, social links, and a companion blog post. All discovered and enriched by AI agents.

6+ distribution channels receiving content from every pipeline run. Two Twitter accounts, two Instagram accounts, YouTube, TikTok, two blogs, and email outreach — all fed from shared content events.

15+ scheduled tasks firing daily or on weekly cadences. Content research, social fetching, carousel rendering, entity discovery, outreach scanning, product marketing, blog publishing.

3 daily content research runs producing blog drafts and hero images every morning before anyone sits down.

20+ reusable automation skills covering the full content lifecycle from research to publishing.

Server-side functions running on Google Cloud that operate independently of any laptop or desktop — tweet fetching, content scoring, webhook processing, email delivery.

The system has been running in production for months. We did not build this as a demo — we built it because we needed it. The content output scales without headcount. The quality stays consistent because every piece routes through human review. The cost is a fraction of what a content team would cost.

A few lessons from running an AI content pipeline in production every day:

Start with one pipeline, not ten. We began with topic research and blog drafting. Once that was reliable, we added social distribution. Then image generation. Then entity discovery. Then outreach. Each pipeline was proven before we added the next. Trying to build the whole system at once would have produced a fragile mess.

Human-in-the-loop is not optional. AI writes good first drafts and mediocre final drafts. The review step is where quality control happens. We tried fully autonomous publishing once. The output was fine 80% of the time and embarrassing 20% of the time. That 20% is enough to damage trust. Every piece of content gets reviewed before it goes live.

Scheduled tasks beat manual triggers. When you have to remember to run something, you forget. When it runs on a schedule, it runs. Our daily content output tripled when we moved from manual skill execution to scheduled tasks. The system assumes the human is busy and acts accordingly.

The knowledge graph matters more than the AI model. The quality of output depends more on the context the AI has than on which model you use. Our Obsidian vault — with CRM contacts, strategy documents, brand voice guidelines, content history, and operational state — is what makes the output feel informed rather than generic. Garbage context in, garbage content out, regardless of how smart the model is.

Composability beats complexity. Twenty small skills that each do one thing well are more reliable and maintainable than one massive workflow that tries to do everything. When something breaks — and things break — you fix one skill, not the whole system.

Everything described in this post is the exact system we offer to build for other businesses through Early to AI. Not a simplified version. Not a template. The same architecture — content discovery, blog drafting, visual generation, multi-channel distribution, human review dashboard — tuned for a different industry.

The methodology is the same regardless of domain:

1. Shadow. We observe your current content workflow. Where do you spend time? What is repetitive? What requires your expertise and what is just execution?

2. Systematize. We map every step and separate the decisions from the mechanical work. Topic selection might be a decision. Reformatting a blog post into a tweet is execution. The execution gets automated. The decisions get surfaced for your review.

3. Ship. We build the pipeline — skills, scheduled tasks, integrations, review dashboard — and go live. Your role shifts from content creator to content editor. The system does the first 90%. You do the last 10% that requires your voice and judgment.

The system you get in month one is good. The system you have in month six is meaningfully better, because every edit you make and every approval you give teaches the pipeline about your voice, your standards, and your audience.

We will show you the live dashboards from our own pipeline. Not slides. Not mockups. The actual Firestore collections, the scheduled task logs, the Post-Bridge drafts, the content output. Then we will have a conversation about what yours looks like.