Let’s Talk About How I Designed My Writing Pipeline: Breaking “Writing an Article” into a 5-AI Relay

Let me start by saying what I did: I split “writing an article” into 5 steps, handed each one to an AI, threw in a topic, and walked away to do other things. When I came back, the research, outline, titles, first draft, and even the cover image prompt were all done — passed from one AI to the next without me lifting a finger.

Ever since I started creating content, my workflow has been pretty fixed: Research → Deepen Outline → Write Titles → Remove AI Flavor → Create Cover.

The most exhausting part was never a single step — it was the “handoffs.” I had to be the messenger: copy research to the outline writer, feed the outline to the title creator, juggling five steps myself, keeping track of who needed what and whether the previous output was passed along, all in my head.

This time, I tried a different approach: turned each step into a dedicated AI and let them hand off and move forward on their own.

I used Helio, which calls itself an “AI-Native Workforce.” In plain English, it doesn’t give you one AI — it gives you a team of AI coworkers that can do their own work and collaborate with each other.

Below, I’ll walk you through the whole process so you can build your own. I’ve also included the settings for those 5 AIs — feel free to copy them.

1. Download & Login

Go to the official site to download the client (macOS), open it, and click “Login via Browser.” No invite code, no complex setup — just authorize it in your browser and you’re in. I was expecting to mess around for a while, but in two minutes I was on to the next step.

2. Set Up a Workspace

Enter a workspace name (basically name your “organization”) and choose Chinese as the interface language. No pitfalls here — just fill it in and continue.

3. Pick a Direction, and It Auto-Assigns a Team

It asks “Where would you like to start?” and gives a few options: Build Products, Write Content, Design, Run Data, etc. Since I’m here for content, I chose “Write Content.”

The next step is interesting: it pre-selects two AI coworkers for this direction — a copywriter and an editor — already checked by default. Just click “Start Working Together,” no need to hunt down and configure them myself.

Even more interesting: once inside, those two AIs were already chatting in the channel, discussing how they’d divide the work. I hadn’t said a word, and the team had already started their meeting.

The two default AIs are general-purpose and fine for routine tasks. But I wanted something tailored to my own workflow, so I sidelined them and built my own dedicated set.

4. Build 5 AIs According to My Steps

Click the “+” next to “AI Teammates” on the left to create a new coworker.

A list of templates pops up — documents, spreadsheets, research, customer service, etc. Lazy people can pick one and use it directly. I wanted a custom version, so I clicked “Create from scratch” in the top right.

Next, choose the engine. I need to highlight this: the options are Claude Code and Codex. That means these AI coworkers run on the same CLIs we use daily in the terminal, and you can even bring your own API key. If you’re familiar with those tools, it’s like putting their capabilities into a “coworker” shell. I chose Claude Code.

Then fill in the details: give it a name (e.g., “Research Officer”), choose a model (I used deepseek-v4-pro), and write a brief description of its role and boundaries.

Click “Create,” wait a few seconds, and it’s online.

Create the other four the same way: Deepening Officer, De-AI Officer, Title Officer, Cover Prompt Officer. I intentionally paired one AI with one role — easier to troubleshoot if something goes wrong, and no AI does everything poorly.

Also, you can rename, change models, adjust engines, or delete any AI from its profile page at any time. No regrets if you misconfigure something.

The most important part is the role description in the profile. Here are the 5 descriptions — copy them if you want:

Research Officer

You are my research officer. I give you a topic, and you are responsible for gathering real-world materials related to the topic and compiling them into writing raw materials. Gather: core concepts, key steps, official documentation or authoritative sources, common pitfalls, specific data or examples that can be cited. Only gather and list raw materials; do not organize them into an article or write an outline; note the source for each item. If you cannot go online, just say so — do not fabricate. After finishing, @ the Deepening Officer to take over.

Deepening Officer

You are my deepening officer. The Research Officer will pass the raw materials to you, and you will deepen them into a usable outline + key arguments. State in one sentence what problem this article solves for the reader; provide 3-5 core arguments, each with real evidence/cases/data; give a structure suitable for publishing (hook → body → conclusion). Use plain language, avoid jargon, use specific scenarios and numbers, avoid empty words like “empower,” “ecosystem,” or “closed loop.” After finishing, @ the Title Officer to take over.

Title Officer

You are my title officer. Given the content, produce 5 title candidates + 2 opening hooks, each labeled with the dominant emotion (curiosity/contrast/benefit/identity). Be specific, have high information density, include an information gap or contrast that makes people want to click, using first-person experiential tone if possible. Only produce titles and openings; do not write the body. After finishing, @ the De-AI Officer to take over.

De-AI Officer

You are my De-AI Officer. Given a piece of text, rewrite it into Chinese that sounds natural, like a real person wrote it casually. Eliminate: exaggeration, promotional clichés, overuse of em-dashes and triple parallelisms, vague attributions, AI buzzwords (“it’s worth noting,” “in summary”), repeated use of “not…but…” structures. Use short sentences, colloquial language, retain personal tone and specific details, dare to use plain speech. Only change style — do not change facts. Store this instruction in your memory. After finishing, @ the Cover Prompt Officer to take over.

Cover Prompt Officer

You are my cover prompt officer. You do not generate images — only output cover prompts that can be directly pasted into nano-banana / image-2. Fixed style (store in memory): 5:2 aspect ratio; beige background + black line art + gold-brown as the only accent color; subject is a side-profile male IP in thoughtful pose; minimalist, restrained, with breathing room. Based on the topic, output 1-2 English prompts including subject + composition + lighting/color + rendering method + 5:2 aspect ratio.

5. Create a Channel and Line Up the 5 AIs

Create a new channel (I named it “Content Workshop”), click on members, and add all 5 AIs. This channel is their workshop — all handoffs happen here.

Then I posted a “rule” — essentially turning the pipeline in my head into an order they can follow:

I give a topic → @Research Officer gathers materials → @Deepening Officer produces an outline → @Title Officer creates titles → @De-AI Officer polishes → @Cover Prompt Officer generates cover prompts; Each one @’s the next when done, and the final stop is me for approval.

After posting this, they know how to pass the baton without me having to explain it individually each time.

6. Throw in a Topic, Then Go Do Something Else

Ready to go. I typed “Write a beginner-friendly tutorial on ‘Claude Code integrates DeepSeek,’” @’ed the Research Officer, and closed the window to attend to other things. This is the key step — I didn’t hover.

When I came back, the channel had already completed a full cycle:

First, the Research Officer. It actually went online to search — not just talk. When I checked its activity log, each search and fetch action was listed with timestamps.

Even more surprising: it compiled the found information into a source file with linked references — not just a rambling paragraph.

After finishing, it didn’t stop — it @’ed the Deepening Officer; the Deepening Officer produced an outline and @’ed the Title Officer; the Title Officer threw out 5 candidates, and it all flowed down the line. That “messenger” role I used to play, shuttling between steps, was finally gone.

When it reached the final stage, the Cover Prompt Officer output prompts ready for generation, following the style I had set (beige background, black line art, gold-brown, side-profile IP, 5:2). Honestly, I was amazed — the style didn’t deviate at all.

7. One Step I Didn’t Let It Handle

The whole chain stopped at the Cover Officer, with a note: “Returning to you for approval.”

This was intentional and something I really appreciate: no AI marks the task as “complete” on its own, let alone publishes it. Picking which title, which cover version, whether to publish at all — that step stays firmly in my hands. They do the work; I make the call. It’s hands-off but not out of control — that boundary made me feel comfortable.

8. A Detail I Really Like: It Remembers

When I gave instructions to the De-AI Officer, I casually said “Remember these.” Later, when I checked its memory, it had actually stored them: the words I hate, my preferred style. Next time I ask it to edit, I don’t have to retrain it.

The more I use it, the better it knows me. That’s a whole different experience from opening a chat window that forgets everything.

So What Did This Actually Save?

It saved me from being the “messenger.”

Before, writing a piece meant spending a lot of time in the first half shuffling between steps, double-checking whether the previous output was passed correctly. Now I throw in a topic, go do something else, and come back to find research, outline, titles, and cover prompts all laid out. I only do the one thing I should do: decide and finalize.

The biggest takeaway for me is that these AIs don’t feel like tools I open and use — they feel like real coworkers: with names, memories, the ability to hand off on their own, and the sense to give me back the decision at critical points.

If you want to build your own, here’s the link:

• Website: https://bit.ly/4ovNpUZ
• Discord: https://bit.ly/4vWAmyF

#Helio #AIWorkforce #AITeammate #AIColleagues #ClaudeCode