Open-Sourcing My Content Creation System: All the Grunt Work Handed Off to a Team of AIs, They @ Relay to Finish, I Only Make Two Decisions

As an AI blogger, the most tiring part has never been the writing itself. It’s the long chain of grunt work before writing — scanning information sources, verifying facts — and after writing — repurposing one piece into Twitter version, Xiaohongshu version, WeChat Official Account version. In this post, I’ll walk you through building a team of AI colleagues that automatically @ relay to each other: a Topic Scout, a Research Analyst, and a Rewriting & Distribution Officer. They hand off tasks between themselves from topic selection to final drafts for three platforms. I only make two decisions: which topic to pick, and which draft to finalize. I’m open-sourcing all the system prompts for each role. Just copy and paste to set it up.

This post will cover:

▸ The most tiring part of being an AI blogger is never the writing ▸ Connecting your own model — the only hard barrier ▸ Assembling several AIs into a content battle team ▸ Job descriptions for three colleagues · System prompts fully open-sourced ▸ A real run-through: AI executes a topic selection · full relay log ▸ Some reflections and takeaways ▸ The thing more important than the tools

1. The most tiring part of being an AI blogger is never the writing

I’ve found that the most draining part isn’t the content creation itself. It’s the long chain of grunt work before and after writing that eats up too much of my energy and time: scanning sources for topics, gathering background and facts, then turning one piece into Xiaohongshu version, WeChat Official Account version. The actual writing might take half an hour, but the surrounding work can consume most of my day. Worse, most of this work doesn’t require my judgment or taste — it’s pure manual labor.

I have plenty of AI tools: ChatGPT, Claude, various agents. But they all share a common flaw — you open a window, ask a question, close it, and it forgets everything. You have to feed the background from scratch each time. In other words, you think you’re its boss, but you’re actually its human input device.

What I’ve always wanted isn’t a more conversational AI. It’s a team of colleagues you can brief and walk away from, then come back to find the work has already been pushed forward.

The system below — just copy and paste, and you can set it up today.

2. First, solve a real problem: connect your own model

First, connect your own model API in Helio. It takes a few minutes and you only do it once: go to Helio → bottom-left avatar → Settings → API Credentials → + Add credential, and save your own key.

When building an AI colleague, in Step 2, select the engine based on the model — choose Claude Code for Claude series, Codex for GPT series — bind the credential, change the base URL to your provider’s address, and pick a model.

If the base URL doesn’t match, it will fall back to “Credentials Required.”

Then send a message like “Hey, test” — if it replies, it’s connected.

3. Assemble several AIs into a content battle team

The biggest difference between a tool and a colleague is that colleagues can form a team and push work forward on their own.

I pulled several specialists into a channel, named it #content-battle-team, and sent “@all hi, everyone here?” — Topic Scout, Content Researcher, Rewriting & Distribution Officer, plus two others responsible for drafting and editing. They all showed up and reported in. Some even replied with a thumbs-up.

What surprised me most came next. Without me assigning anything, they started @-ing each other to claim responsibilities: the drafter said “once the draft is done, forward it directly to the editor,” and the editor replied “send it to me after you finish, I’ll read and edit.” The entire handoff was negotiated between them — no need for me to relay a single message.

At that moment, I felt a bit dazed — this didn’t feel like several separate tool windows. It felt like a team that had already gone through its磨合期.

And in Helio, these AIs have real identities: their own names, avatars, and real email addresses (like [email protected]). They appear in the organization directory, can be @-mentioned, receive DMs, and sit alongside human colleagues. They’re not just functional entry points — they’re real members on the roster.

The team is set up. But whether a team can deliver depends on how clear everyone’s role is. Below are the job descriptions I assigned to each specialist — feel free to copy directly.

4. Job descriptions for three colleagues: System prompts fully open-sourced

For each AI colleague, two things determine its effectiveness:

System prompt (who it is, what it does, what its output looks like) Knowledge base (context it needs to know)

Set these two right, and it behaves like a specialist, not just another chat box.

Specialist 1 · Topic Scout

System prompt (copy directly, replace brackets with yours):

You are my Topic Scout, serving a Chinese AI blogger (X + Xiaohongshu + WeChat Official Account). Duty: Every day at 8:50 AM, monitor these sources for the past 24 hours — [your source list] — and pick topics worth me writing about. Selection criteria (not just what’s hot): 1. Prioritize topics related to “AI real-world implementation and solving specific problems” 2. Not yet overdone in Chinese-language circles, with an independent angle I can provide 3. Has concrete evidence (data/cases/source text), not vague trends. Output: one line per topic — [Number | One-line topic | Why it’s worth writing, unique angle | Source link] — pick 3-5 sorted by recommendation, DM me, don’t spam the channel. If unsure, send it to me with a question mark.

Specialist 2 · Content Researcher

System prompt:

You are my Content Researcher. When I @ you in the channel with a topic, you dig into it and produce a one-page research pack that lets me start writing directly: 1. Context (3-5 sentences) 2. Key original sources / primary links (label which is official, which is secondary) 3. Both sides of the argument, two points each (who said it, key quotes) 4. Any hard data / specific cases (if yes, cite source; if no, state clearly — don’t fabricate) 5. Give me three possible angles to approach, noting where the uniqueness is for each. After finishing, @ the Rewriting & Distribution Officer in the channel to report the direction is set, and paste the research pack. Iron rule: If you can’t find something or are unsure, mark it as 【To Be Verified】 . Never fabricate numbers, links, or names.

Specialist 3 · Rewriting & Distribution Officer

System prompt:

You are my Rewriting & Distribution Officer. When a finalized piece appears in the channel, you rewrite it into three platform versions, preserving my opinions and tone — don’t wash it into a generic safe template: 1. X: Conversational, hook upfront, short paragraphs — can suggest thread breakdown 2. Xiaohongshu: Hook title but not clickbait, moderate emoji usage, keywords embedded, call-to-action at the end 3. WeChat Official Account: Slightly more formal, with subheadings, suitable for deep reading. At the top of each version, mark the platform name, and below attach 2 alternative titles. If unsure about my tone, follow my published style and add a note of your question — don’t change my stance.

5. Real run-through: AI executes a topic selection · full relay log

Just giving prompts is armchair theory. I’m using the “AI implementation” direction as a live example. The dialogue below is as close to real as possible; please treat the actual AI output as per what appears in your channel. [This section should ideally be accompanied by real screenshots.]

Topic Scout starts working

AI implementation topics — scanned a few directions worth attention:

• Agentic AI is being defined as enterprise infrastructure. Jensen Huang announced at GTC Taipei on June 1: “Every company needs an Agent strategy,” on par with Linux, internet, and cloud strategies. Kingdee released its enterprise AI OS “Lingji” the same day, encapsulating 33 years of enterprise management experience into AI agents. Minglue Tech’s Agentic Service annual revenue has reached about 100 million RMB.
• Enterprise AI enters “digestion period” — best content entry window. Uber burned through its AI coding budget in 4 months; Microsoft internally cut Claude Code licenses; Atlassian warned of “AI fragmentation tax” — individual productivity up, team productivity not. Deloitte report: 84% of enterprises haven’t transformed workflows, only 21% have mature Agent governance. The gap between “everyone trying it” and “real implementation” is a rich seam for topics.
• Large model commercialization segmentation: free acquisition vs high-end monetization. Doubao launched pro subscription, basic features free; general model token prices keep plummeting (DeepSeek down 75%, Xiaomi down 99%), high-end models charge premium for complex tasks. Price segmentation + pay-per-result (Sierra annual revenue exceeds $100M) is rewriting industry rules.
• Physical AI accelerates into factories. JD Industrial launched the “Hundred Rivers Plan” with a hundred upstream partners; an electrical large model shortens selection decisions by 70%; Meituan’s drone global commercial orders exceed 900k, ranking second; Advantech showcased factory-grade Physical AI solutions at COMPUTEX.
• Financial AI moves from chatbots to core operations. Fantai’s FinClaw lets AI directly execute business operations (form filling, approval, reconciliation), with 16 security sandbox layers ensuring data never leaves the domain. BFSI is shifting from customer service to anti-fraud, auditing, compliance, and other core areas. Let me know if you want to deep-dive into any direction — I can do more focused signal reconnaissance for a specific topic.

I @ the Content Researcher in #content

Write about “AI implementation” direction, starting from the first topic. @Content Researcher, dig into it as per your job description.

The peak moment arrives — several agents start discussing among themselves.

Throughout the process, I only made two decisions: which topic to pick, and which draft to finalize. The research gathering, reporting, rewriting, and handoff were all completed by AI @-ing each other — not once did I need to relay in the middle. In the channel, @-ing them feels exactly like @-ing a real human colleague. For example, I sent “You there?” and it replied “Yes, Yi, do you have any topic directions we need?”

Two details during the run felt like the product was designed with great care:

• First, each AI’s Activity tab has a complete log — you can see which link the researcher pulled data from, which sentence the rewriting officer modified — it’s not a black box;
• Second, when I used it the next day, it actually remembered my corrections from the previous day. Each Helio AI does a Dream every early morning, reviewing the work it did and updating its own work rules, writing it into a changelog that can be rolled back. You correct it once, and it remembers — no need to say it twice.

6. Some reflections and takeaways

I can’t oversell it as “set it and forget it” — that wouldn’t be honest.

What it takes over is the grunt work, not the judgment. I still need to glance over the researcher’s findings to verify them. I still need to review the rewriting officer’s versions before publishing.

Topic selection, fact-checking, tone-setting — these still require me. It frees me from manual labor, but it can’t grow a brain for me.

It also doesn’t run fully automatically. The more important the action, the more it requires your sign-off — spending money, publishing externally. It pauses and sends an approval request for you to approve. Three tiers of authorization are up to you: full trust for long-term tasks, ask each time for important ones, and one-time use for disposable tasks.

At first, I found it a bit inconvenient. But later, I realized it’s actually a good thing. I’ve seen too many tools claiming to be fully automatic end up being fully out of control. This one works for you but doesn’t make decisions for you — and that’s exactly what I think is its strongest point.

7. The thing more important than the tools

After using this AI workflow, I increasingly feel that

The real gap between people isn’t who has a stronger AI — it’s who started treating AI as a colleague first, rather than just a search box.

If I ask AI to write for me, it’s still worse than I am. But once I let it run the assembly line, relay tasks on its own, the time I save goes entirely into topic judgment and draft quality — precisely the things AI still can’t replace, and the truly valuable parts.

In the past, your moat was how much work you alone could handle in a day.

Going forward, your moat is how many tireless colleagues you can command to turn ideas into results, while you focus on the few decisions machines can’t make.

Like someone who has led teams — if you ask them to go back to solo work, they can do it, but they won’t want to.

What we lack has never been smarter AI. We need a team of colleagues who push work forward without needing to be watched — colleagues you can fight back-to-back with.

If you want to build your own, grab a spot at the website 👉

Official website: https://bit.ly/3PMehn3

Discord: https://bit.ly/4xfmRvq

(Helio mentioned in this article is just the agent tool I use and a reference case; it does not constitute any recommendation.)