Key Takeaways
- AI assists, humans decide: The pipeline handles the time-consuming mechanical work — monitoring, research, drafting — while the editor reviews and approves every piece before it goes anywhere.
- OpenClaw + NemoClaw on DGX Spark: The stack runs entirely locally, keeping content drafts and research data on-device with no cloud dependency.
- Bilingual from day one: Separate specialist agents produce English and Traditional Chinese drafts simultaneously, each tailored to its audience — not machine-translated.
The Content Problem Every Niche Site Faces
Keeping a specialist content site consistently updated is genuinely hard. In fast-moving sectors — finance, technology, markets — relevant stories break daily, and staying on top of them while also writing clearly, accurately, and for multiple audiences is a significant time investment. The usual options are: hire writers (expensive), write everything yourself (unsustainable), or publish less (hurts SEO and reader engagement).
The fourth option — one that most site operators have not yet explored properly — is to use AI agents as content assistants rather than content replacements. Not agents that publish automatically and unsupervised, but agents that do the research, monitoring, and first-draft work so that a human editor can focus entirely on reviewing and refining rather than starting from a blank page every day.
This guide walks through how we built exactly that at SavingsAI, using OpenClaw and NVIDIA NemoClaw running on a local Nvidia DGX Spark.
The Core Principle: AI Drafts, Humans Decide
Before getting into the technical setup, it is worth being explicit about the design philosophy — because getting this right is what separates a useful content pipeline from a problematic one.
The non-negotiable rule
No content leaves the pipeline without a human reading, approving, and consciously choosing to publish it. The AI agents are researchers and drafters. The editor is the gatekeeper. This is not just good practice — it is what produces content worth reading, and it is what keeps the pipeline trustworthy over time.
AI language models are extremely good at structured drafting tasks: summarising source material, writing to a template, adapting tone for different audiences, and producing consistent output at scale. They are less reliable at editorial judgement: deciding what is genuinely important today, what angle serves the reader, and what context a specific audience needs. The pipeline is designed to let each side do what it does best.
The Five-Agent Architecture
The pipeline uses five specialised OpenClaw agents, each with a single clearly defined responsibility. They communicate via a shared state file rather than calling each other directly, which keeps each agent simple, independently testable, and easy to replace or upgrade.
Each agent runs as a separate OpenClaw process and connects to a dedicated thread in a private Telegram group — the editor's control interface. This means the entire pipeline is managed through a chat app the editor already uses, with no new UI to learn.
| Agent | Role | What it does |
|---|---|---|
| Monitor | Orchestration | Starts the daily pipeline, resets shared state, relays progress from all agents to the editor's main Telegram thread in real time |
| Researcher | News & source monitoring | Fetches today's relevant articles from configured feeds, ranks by topic relevance and recency, presents a clean summary table for the editor to select from |
| Content Creator | Bilingual drafting | Produces two drafts per selected article — one in English, one in Traditional Chinese — each tailored to its audience with appropriate tone, angle, and context |
| Publisher | Formatted output | Takes approved drafts and formats them into the correct output structure — ready for the editor to paste, schedule, or hand off |
| QA Reviewer | Quality checking | Reviews published content against the approved draft — checks for accuracy, formatting, and completeness; flags any discrepancies for the editor |
The Workflow: Step by Step
The pipeline runs on a daily schedule triggered by OpenClaw's built-in cron system. The editor interacts at three deliberate checkpoints — everything else runs autonomously between those gates.
Daily trigger Automatic
The Monitor agent starts automatically each morning, resets the pipeline state, and kicks off the Researcher. The editor receives a "Pipeline starting" notification in Telegram.
News monitoring & ranking Automatic
The Researcher fetches today's articles from configured sources, filters to the relevant topic area, ranks by relevance and importance, and posts a formatted summary table to Telegram — typically the top 10–15 stories of the day.
Editor selects stories Human checkpoint
The editor reviews the table and replies with the numbers of the stories worth covering. This is pure editorial judgement — the AI has done the legwork of finding and ranking; the human decides what is genuinely worth the audience's attention today.
Bilingual drafting Automatic
The Content Creator agent produces two drafts per selected story — one in English, one in Traditional Chinese. Each follows a structured template: lead paragraph, key facts, context for the target audience, and a summary. All drafts are posted to Telegram for review.
Editor reviews & approves drafts Human checkpoint
The editor reads every draft. They can approve outright, request specific edits ("rewrite the lead", "add context about ISA rules"), or reject. The Content Creator rewrites to order and resubmits. Nothing proceeds until the editor is satisfied.
Formatting & output Automatic
The Publisher agent formats each approved draft into the correct output structure — clean HTML, structured metadata, or plain text depending on how the pipeline is configured. Output is saved and flagged as ready.
Final publish decision Human checkpoint
The editor makes the final call on what gets published, when, and where. The pipeline produces ready-to-use content; the decision to actually put it in front of readers is always human.
QA review Automatic
The QA Reviewer checks published content against the approved draft — flagging any formatting issues, missing elements, or discrepancies. Any problems are reported to the editor immediately.
Why Telegram as the Interface?
One of the underappreciated advantages of the OpenClaw architecture is that agents live natively in chat apps rather than requiring a dedicated dashboard. For a content pipeline, this is particularly useful — the editor already has Telegram open, receives the pipeline notifications alongside their other messages, and can approve or request changes in seconds without switching context.
The pipeline waits indefinitely at each human checkpoint. There is no timeout, no pressure to respond immediately. The editor can review drafts carefully during their normal working day rather than being constrained by a processing window.
Bilingual Content: Not Translation, Audience Adaptation
A key design decision in our pipeline is that English and Traditional Chinese content is not translation — it is separate drafting for separate audiences. The Content Creator agent has different system prompts and context for each language output:
- English drafts are written for UK readers — the context, comparisons, and implications are framed around UK savings rates, UK regulatory environment, and UK financial behaviour.
- Traditional Chinese drafts are written for Hong Kong investors and diaspora following UK markets — the same story is framed differently, with context about how UK changes compare to HK financial norms and why the news is relevant to that specific audience.
This approach produces significantly better content than translation and is one of the things AI agents handle particularly well — adapting the same underlying facts for different audience contexts at no extra time cost.
The Technology Stack
What runs this pipeline
What the Editor Gains
The practical benefit of this setup is a significant reduction in the time and cognitive load required to produce consistent, quality content. The tasks that consume the most time without requiring editorial judgement — scanning feeds, identifying relevant stories, producing structured first drafts — are handled automatically. The editor focuses entirely on the tasks that require human judgement: deciding what matters, ensuring accuracy, refining tone, and approving publication.
For a single person running a specialist content site, this is the difference between publishing one or two pieces a week (what is sustainable writing from scratch) and publishing three to five per day (what is sustainable with AI drafting assistance). The quality ceiling is still set by the editor — the AI accelerates the volume without compromising the standard.
A note on AI content and AdSense / search quality
Google's guidance on AI-assisted content is clear: content produced with AI assistance is acceptable as long as it is original, provides genuine value to readers, and has meaningful human oversight. A pipeline with three human approval gates — story selection, draft approval, and publish decision — satisfies this requirement comfortably. The AI is a tool; the editor is the author.
Getting Started with This Stack
If you want to build a similar pipeline for your own content site, the starting point is OpenClaw:
- Install OpenClaw on your local machine — it runs on Mac, Windows, or Linux and works with Claude, OpenAI, or local models via Ollama
- Connect it to Telegram (or WhatsApp/Slack if you prefer) and create a basic single-agent setup
- Build out a Researcher agent first — an agent that monitors your topic feeds and presents a daily briefing is useful immediately, even before any content drafting
- Add the Content Creator agent once the monitoring pipeline is stable
- Add NemoClaw on top if you are running NVIDIA hardware and want the additional guardrails and local model performance
The full stack runs on a DGX Spark for us, but OpenClaw itself runs on hardware as modest as a Mac mini or Raspberry Pi. Start simple, validate the workflow with your own content type, and scale the infrastructure as needed.
See OpenClaw and NemoClaw Explained
New to these tools? Read our clear overview of what each one does, how they differ, and when you need the full NemoClaw stack versus OpenClaw alone.
Read the OpenClaw + NemoClaw Guide →