Open-Sourcing My Content Creation System: Offload All Grunt Work to an AI Squad That @-Mentions Its Way Through the Pipeline – I Only Make Two Decisions
Being an AI blogger – the most exhausting part has never been the writing. It's the grind *before* you write: scrolling through sources, fact-checking. And the grind *after* you write: turning one piece into a Twitter version, a Xiaohongshu version, a WeChat Official Account version – that whole chain of busywork. In this post I'll walk you step-by-step through building a team of AI colleagues that @-mention each other to pass the baton – a Topic Scout, a Research Analyst, and a Rewriter & Distributor. From topic selection to final drafts for three platforms, they hand off to one another. I only make two calls: which topic to pick, and which draft to approve. I'm open-sourcing every system prompt for each role – copy-paste and you're set.
This post will walk you through:
▸ The most tiring part of being an AI blogger is *never* writing content
▸ Connect your own model – the one real hurdle
▸ Assemble a squad of AIs into a content team
▸ Three job descriptions – full system prompts open-sourced
▸ A real run-through: an AI adoption topic tracked from start to finish, relay-style
▸ Retrospective & reflections
▸ The one thing more important than the tool
1. The most tiring part of being an AI blogger is never writing content
After doing this for a while, I've realized the real energy drain isn't writing itself – it's that huge string of busywork *before* and *after* writing. Scrolling feeds for topics, verifying background and facts, then taking one piece and turning it into a Xiaohongshu version, Official Account version. The actual writing might take half an hour, but the before-and-after can eat up most of a day. What makes it worse: most of this work doesn't need my judgment or taste – it's pure manual labor.
I have plenty of AI tools – ChatGPT, Claude, various agents – but they all share a flaw: you open a window, ask a question, close it – and it forgets everything. Next time you have to feed it the context from scratch. You think you're the boss, but really you're its human input device.
What I want is never just a more talkative AI. I want colleagues I can assign a task to and walk away – come back and find the work has already moved forward.
Below is the system – copy it and you can set it up today.
2. First, solve the real problem: connect your own model
In Helio, connect your own model API. Takes just a few minutes, and you only do it once: Go to Helio → Avatar in bottom left → Settings → API Credentials → + Add credential. Store your own key.
When you create an AI colleague, in Step 2, choose the engine by model – Claude series → Claude Code, GPT series → Codex – bind the credential, change the base URL to your provider's address, then pick a model.
If the base URL doesn't match, it'll fall back to "Credentials required."
Then send a "Hello" to test – if it replies, you're good.
3. Assemble a team of AIs into a content squad
The biggest difference between tools and colleagues: colleagues can form a team and push work forward on their own.
I added several specialists into one channel, named it "#ayi-content-squad", and sent "@all hey everyone, you here?" The Topic Scout, Content Researcher, Rewriter & Distributor, plus two handles for drafting and editing – one by one they all checked in, some even replied with a thumbs-up.
What surprised me most came next. Without me assigning anything, they started @-mentioning each other to divide the work. The first-draft writer said "when the draft is ready, pass it to the editor," and the editor replied "after you finish, throw it my way, I'll read and edit." The whole handoff was negotiated between them – I didn't have to relay a single message.
For a moment, I was a bit stunned. It didn't feel like several separate tool windows – it felt like a team that had already gelled.
And these AIs in Helio have real identities: names, avatars, an actual email (like alice@yourcompany.helio.im), visible in the org directory, can be @-mentioned and receive DMs, sitting right next to human colleagues. Not just a function endpoint – real members on the roster.
Squad assembled, but whether it can deliver depends on whether each role is clearly defined. Below are the job descriptions I gave each specialist – you can copy them directly.
4. Job descriptions for three colleagues – system prompts fully open-sourced
For every AI colleague, two things really determine its usefulness:
- System prompt (who it is, what it does, what its output should look like)
If these two are right, it acts like a specialist – not just another chat window.
Specialist 1 · Topic Scout
System prompt (copy-paste, replace the brackets with yours):
You are my Topic Scout, serving a Chinese blogger focused on AI (X + Xiaohongshu + WeChat Official Account).
Duty: Every day at 8:50 AM, monitor these sources for the last 24 hours – [your source list] – and pick out topics worth my writing.
Criteria (filter by these, not just hotness):
① Priority to topics related to "AI real-world adoption, solving specific problems"
② Not yet over-covered in Chinese circles, with an independent angle I can add
③ Backed by concrete evidence (data / case / original link), not vague trends
Output: One line per item – [Number | One-line topic | Why it's worth my writing, unique angle | Source link]. Pick 3-5 items ranked by recommendation, DM them to me – don't spam the channel. When unsure, rather show me – mark it with a question mark.
Specialist 2 · Research Analyst
System prompt:
You are my Research Analyst. When I @-mention you in the channel with a topic, you dig it up and produce a one-page research pack that lets me start writing directly:
① Background and context (3-5 sentences)
② Key original sources / primary links (clearly mark which are official, which are secondary)
③ Two supporting and two opposing voices each (who said it, key points from original)
④ Any hard data / concrete cases (mark the source if yes; if no, say so – don't fabricate)
⑤ Three possible angles for me to approach the topic, each with what makes it unique
After finishing, @-mention the Rewriter & Distributor in the channel to confirm the direction is set, and paste the research pack.
Iron rule: If you can't verify or are uncertain, mark it as [TO VERIFY] clearly. Never make up numbers, links, or names.
Specialist 3 · Rewriter & Distributor
System prompt:
You are my Rewriter & Distributor. When a finalized piece of content appears in the channel, you rewrite it into versions for three platforms – retaining my opinion and tone, not laundering it into a bland safe template:
① X (Twitter): conversational, hook upfront, short paragraphs – can give thread split suggestions
② Xiaohongshu: title with hook but not clickbait, moderate use of emoji, keywords well-placed, end with an interactive hook
③ WeChat Official Account: slightly more formal, with subheadings, suitable for deep reading
At the top of each version, label the platform name, then attach 2 alternative titles below.
When unsure about my tone, follow the style of my published posts, and note your question next to it – don't change my stance.
5. Real test: Running an "AI adoption" topic – full relay record
Just giving prompts is theory. I'll take the "AI adoption" direction. The conversation below is as real as I can capture – the exact AI output will depend on what you see in your channel [suggest adding actual screenshots].
Topic Scout gets to work
AI adoption topics – I've spotted these directions worth attention:
* Agentic AI pegged as enterprise infrastructure. Jensen Huang at GTC Taipei June 1 declared "Every company needs an Agent strategy" – on par with Linux, Internet, cloud strategies. Kingdee launched "LingJi," an enterprise AI OS, six-layer architecture packaging 33 years of enterprise management experience into AI agents. MinglyTech's Agentic Service annual revenue ~100 million RMB.
* Enterprise AI enters a "digestion phase" – the best window for content. Uber's AI coding budget burned through in 4 months; Microsoft internally cut Claude Code licenses; Atlassian warns about "AI fragmentation tax" – individual productivity up, team productivity flat. Deloitte report: 84% of enterprises haven't transformed workflows, only 21% have mature agent governance. The gap from "everyone tries it" to "real adoption" is a topic goldmine.
* LLM commercialization tiers: free acquisition vs. high-end monetization. Doubao launched Pro paid subscription, basic features free. General model token prices continue to crash (DeepSeek down 75%, Xiaomi down 99%). High-end models leverage complex task capabilities for premium pricing. Tiered pricing + pay-per-outcome (Sierra's annual revenue exceeded $100M) rewriting industry rules.
* Physical AI accelerating into factories. JD Industrial joined with 100+ upstream companies to launch the "Baichuan Plan," electrical LLM cuts selection decisions by 70%. Meituan's drone global commercial orders rank second, over 900k. ADLINK showcases factory-grade Physical AI at COMPUTEX.
* Financial AI moving from chat to core operations. FinClaw by FanGeTech lets AI directly execute business ops (fill forms, approvals, reconciliation), 16-layer security sandbox keeps data on-prem. BFSI is penetrating from customer service to fraud detection, auditing, compliance. Let me know if you want deeper signal on any of these directions.
I @-mention Content Researcher in #content
Write about "AI adoption" direction. Cut from point 1. @ResearchAnalyst, dig it up per your job description.
Here comes the peak – then the agents start discussing among themselves.
Throughout the whole process, I only made two decisions: which topic to pick, and which draft to finalize. Everything in between – research gathering, reporting, rewriting, passing – the AIs handled with @-mentions among themselves. I never had to relay a single message. @-mentioning them in the channel is exactly like @-mentioning a human colleague – for example, when I logged in and sent "You there?", one replied "Yes, Yi, any topic direction you need our help with?"
Two details during the run made me feel this product is really thoughtfully designed:
- Every AI has an Activity tab with a full log – which links the researcher pulled, which sentences the rewriter changed, all traceable, not a black box.
- The next day when I used it again, it actually remembered my corrections from the day before – each Helio AI does a "Dream" every day at midnight, reviewing the day's work, updating its own work rules, writing a changelog with rollback ability. Correct it once, it remembers on its own – you don't need to say it twice.
6. Retrospective & reflections
I can't sell this as "set it and forget it, money rolls in" – that wouldn't be honest.
What it took over is the grunt work – it doesn't touch judgment. I still need to glance at the researcher's pack to verify. The rewrite versions still need my review before I publish.
Topic taste, fact-checking, tone setting – I still have to do these myself. It freed me from manual labor, but it can't grow a brain for me.
It also doesn't run fully autonomous. The more important the operation, the more it requires your approval – spending money, posting content publicly – it pauses and sends an approval request for your sign-off. Three tiers of authorization: long-trusted roles let go, important ones ask every time, one-time tokens burn after use.
At first I found it a bit annoying, but later I thought – actually it's good. I've seen too many tools that claim to be fully automatic turn into fully out of control. This tool works for you but doesn't make decisions for you – I think that's actually its best feature.
7. The one thing more important than the tool
Using this AI workflow more and more, I've come to feel:
The real difference between people isn't whose AI is more powerful – it's who first treats AI as a colleague, not just a search box.
If I let AI write for me, it writes worse than I do. But once I let it run the production line, passing the baton by itself – I can pour all the saved time into topic taste and draft judgment. Those two things are exactly what AI can't replace yet – the truly valuable part.
Your moat used to be how much work one person can shoulder in a day.
Going forward, your moat is how many sleepless colleagues you can command to turn ideas into results – while you focus on the decisions machines can't make.
It's like someone who's led a team for a long time – they could go back to solo work, but they wouldn't want to.
And what we lack isn't a smarter AI. We need a squad of colleagues who'll push work forward without you having to watch over them – the kind you can fight back-to-back with.
Want to build your own? Grab a spot on the official site 👉
Official site: https://bit.ly/3PMehn3
Discord: https://bit.ly/4xfmRvq
(The "Helio" mentioned in this post is just the agent tool I use and a reference case for the article – not a recommendation.)