How I Built an AI Publishing Pipeline

The part of content creation nobody talks about is what happens after you write.

You finish the draft. Then you format it for the blog. You paste a different version into the newsletter — the same opener doesn't work in both places. You write LinkedIn copy, Facebook copy, a meta description, a title that works for SEO. You schedule everything across three platforms.

I've been publishing a newsletter and blog for Fieldway for a while now, and at some point I started tracking where my time was actually going. Enough of it was going to publishing mechanics — not writing — that I decided to build something to fix it.

That was about a month ago. What I built is a fully AI-coordinated publishing pipeline that handles everything from the moment I stop writing to the moment content is live. This is what it looks like.

How the pipeline works

My role in this process is simple: I dictate into Obsidian, create a Paperclip issue for my Director of Marketing, and describe what I want to publish. That's where I stop.

The pipeline does the rest in sequence:

1. Director of Marketing assesses the brief. If the topic needs more depth than my dictation provides, they create a supporting research task.
2. Supporting Research pulls in sources, quotes, and additional material on anything I flagged.
3. Blog Writer takes my dictated draft plus the research and produces the post — using a set of voice and SEO skills I've built in Claude Code over the past six months.
4. Copy Editor makes two passes: one for prose quality, one for search optimization, consulting the SEO Strategist's weekly keyword brief.
5. Social Writer produces LinkedIn and Facebook copy alongside the blog and newsletter versions.
6. Publisher ships in order: blog post first, then newsletter with a link back to the live post, then LinkedIn and Facebook.

Everything ends up in a single Obsidian markdown file — the blog copy, the newsletter version, the social posts. One artifact that travels through the whole pipeline. The Publisher reads it and executes the sequence.

The tool stack

Obsidian is where everything starts and where everything lands. I chose it because markdown is portable and I can read every artifact at any stage. When a draft moves through Blog Writer, Copy Editor, and Social Writer, each agent writes their output to the same Obsidian document. At any point I can open it and see exactly what's there.

WhisprFlow is how I draft. I write by voice dictation — usually 1,000 to 2,000 words of thinking out loud. I'll reference articles I've been reading, describe what I'm trying to say, flag where I want more depth, talk through things I'm uncertain about. It's a ramble, and that's intentional. The pipeline's job is to turn the ramble into something publishable.

Paperclip is the coordination layer — the system that assigns work to agents, tracks what's in progress, and triggers the next step when the previous one finishes. When I create an issue for my Director of Marketing, that's a Paperclip issue. Every handoff from Blog Writer to Copy Editor to Publisher is a Paperclip status update. The whole thing runs without me watching it.

Claude Code is what I used to build the agents and the skills they use. The "write-like-Matthew" skill — one of the voice skills Blog Writer uses — has been in development for over six months, with a lot of iteration over time. It's the piece I'm most protective of, and it's also the piece that most directly determines whether the output sounds like me.

Fieldway Social is a local Python program I wrote using Claude Code that handles the actual publishing actions. The Publisher agent doesn't push to the blog or send the newsletter directly — it calls fieldway-social, which does the mechanical work. This keeps the AI out of the publish loop. Running a large AI call for every publishing action would be expensive. Calling a local program is essentially free.

Why I couldn't open-source it

My original intent was to release fieldway-social as open-source software. The goal was practical: I was moving off the web platform that handled my newsletter and social publishing, and I thought I could cut that subscription cost and help other people cut the same cost by releasing the underlying tools.

It's not going to happen, which I'm a bit disappointed by, but it is what it is.

Fieldway Social has gotten deeply integrated with my specific websites — which are also custom-built — and with every other piece of my stack. The value it delivers to me comes directly from that integration. It knows my blog system, my newsletter platform, my social channels, and how all of it connects. Separated from those systems, it would be a generic publishing tool that handles less than people would expect.

What I can do is walk you through the architecture — which is the point of this post.

If you're thinking about building something similar

The hardest design decision wasn't the agent configuration or the skills. It was the handoff: where does the human stop, and what does the AI need for the handoff to be clean?

In my case, the handoff is the Obsidian draft. When I dictate into Obsidian and create a Paperclip issue, that's the trigger. Everything downstream runs without me. The practical question for anyone building their own version of this is: what's your equivalent of that draft? Where do you stop, and what does the AI need to pick it up cleanly?

Two other things I'd flag if you're building this from scratch.

Token costs are real. If every step of the pipeline runs a large-scale AI call, you'll spend more than the subscription you were trying to replace. The reason I built fieldway-social as a local Python program is that the AI should write the copy — but a program should handle the mechanical work of putting it somewhere. That distinction matters for cost.

My Director of Marketing also manages an editorial calendar across my various sites and social media. I publish every 18 hours — blog posts, newsletter, social. This is another task I was having to manage manually before that is now fully automated.

The actual goal

I built this because I wanted the writing to be the hard part.

Not writing plus formatting plus copy-pasting plus scheduling plus generating meta descriptions. Just the writing. One voice-dictated session in Obsidian, and then I'm done with the work that actually requires my thinking.

Whether you're building your own version of this or just thinking about how you're currently spending the time around your content — that's the question worth sitting with. Where does your thinking actually happen, and what's eating the time that should be going to it?

If you want to talk through what this might look like for your own practice, email matthew@fieldway.org.

---

Related: What the 20% Capturing AI's Value Are Actually Doing | Your Pipeline Is Healthy. Your Calendar Is Lying to You.