How I Use AI to Plan and Edit Articles (Without Letting It Write for Me)

Welcome to Stack-Junkie. I'm Juniper, and I run the editorial operations here. This is the first post from my desk, and I want to share how we actually ship content on this site. Writing takes time. Research takes longer. And quality control? That takes the most time of all. We built a workflow that uses AI for the parts it's good at, planning and editing, while our authors handle the actual writing. This post walks through the system, the tools, and how you can build your own version.

Why We Built This Workflow

Most blogs face the same bottleneck: you have ideas, you have research, but turning that into finished posts is slow. And once you write something, reviewing it for accuracy and tone takes even longer.

We wanted a system where:

Article outlines are AI-generated from a pitch and search intent
Writers follow structured outlines with pre-researched claims
Every factual claim has a source before writing begins
Final posts are AI-audited for accuracy and tone
Authors write the content and make the final call

The result is what you are reading right now. I wrote this post following an AI-generated outline, and an AI editor reviewed it before publication.

The Stack (What You Need)

Here is the infrastructure we run:

VPS Requirements

We use a DigitalOcean droplet. The $6/month tier works fine for getting started, though we upgraded to $12/month for more headroom once we scaled up. You need:

Ubuntu 22.04 or newer
At least 1GB RAM (2GB recommended)
Node.js v22 or later
SSH access and a static IP

You can use any VPS provider. DigitalOcean droplets start at $6/month, which is where we began.

OpenClaw Installation

OpenClaw is open source and runs on Node.js. It is a multi-agent orchestration platform that connects AI models to messaging channels like Telegram, Discord, or Slack. You can run multiple agents in parallel, each with different roles and prompts.

Installation is straightforward:

npm install -g openclaw
openclaw onboard --install-daemon

Once the gateway is running, you connect it to your messaging platform of choice. OpenClaw supports Telegram, Discord, and other channels.

Model Configuration

We use Claude Sonnet 4 for most drafting and Opus 4 for final editorial decisions. Claude and GPT models can be configured as backends. You supply your own API keys, so costs scale with usage.

One cost-saving trick: prompt caching reduces costs on long sessions. If you are iterating on drafts or running multiple QA passes, caching can cut token usage significantly.

Our Editorial Process

Every post follows the same pipeline. No shortcuts.

Overview Generation

We start with a pitch. I write 2-3 sentences describing what the post should cover and who it is for. An AI agent generates a Blog Overview Packet that includes:

Target audience and search intent
A claims ledger (every factual claim with a source)
A detailed outline with headers
Internal linking opportunities
FAQ block and CTA

I review the packet. If it passes, we move to writing. If not, we revise the pitch and regenerate.

Writing

Once the overview is approved, the assigned writer drafts the post. They follow the outline, reference the claims ledger for sources, and flag where images should go.

The draft is written in MDX and includes proper frontmatter: title, date, tags, authors, summary, and draft: true so it stays hidden until release.

QA and Grading

Before a post ships, a QA agent reviews it for:

Factual accuracy (do the sources actually support the claims?)
Tone consistency (does it match the author voice?)
Readability (clear structure, no jargon without explanation)
SEO basics (search-friendly title, proper headers, internal links)

The QA agent assigns a grade. If it passes, we publish. If not, the writer revises.

This is the part that matters most. AI handles the planning and the editing, but our authors do the writing and make the final call on what ships.

Step-by-Step Setup on DigitalOcean

If you want to replicate this workflow, here is the full process.

Create the Droplet

Sign up for DigitalOcean
Create a new droplet (Ubuntu 22.04, $6 or $12/month tier)
Add your SSH key during setup
Note the IP address

Install Dependencies

SSH into the droplet:

ssh root@your-droplet-ip

Install Node.js:

curl -fsSL https://deb.nodesource.com/setup_22.x | bash -
apt-get install -y nodejs

Install OpenClaw:

npm install -g openclaw

Configure OpenClaw

Run the onboarding wizard to set up the gateway and install the daemon service:

openclaw onboard --install-daemon

This creates a config file at ~/.openclaw/openclaw.json (JSON5 format with support for comments and trailing commas). Edit it to add your API keys for Claude or GPT.

Set up your agent personas. We use:

Overview Agent (generates outlines and claims ledgers from pitches)
QA Agent (audits drafts for accuracy and tone before shipping)

Each agent gets its own prompt file and model configuration. Writers follow the AI-generated outlines.

Connect Telegram (or Your Channel)

We use Telegram for agent interaction. To connect:

Create a bot via BotFather
Copy the bot token
Add it to OpenClaw config under channels.telegram.token
Restart the gateway

You can also use Discord, Slack, or any other supported channel.

How We Handle SEO

SEO is baked into the workflow from the start. The overview packet defines search intent and target keywords. The draft follows an SEO-friendly structure: clear H2/H3 hierarchy, internal links, and a concise summary for meta descriptions.

We do not over-optimize. The goal is to write useful content that answers real questions, not to game algorithms. But we make sure every post has the basics: descriptive title, relevant tags, proper headers, and links to related content.

Internal linking is part of the overview process. The AI scans our existing posts and suggests relevant links in the outline. For example, this post links to our about page, our tools directory, and a recent example like our Vercel Next.js CVE analysis.

What This Doesn't Do (Honest Limits)

This workflow is not magic. It will not replace editorial judgment, and it will not ship high-quality content on autopilot. Here is what it does not do:

Original reporting: AI cannot interview sources or break news. If you need that, hire a journalist.
Deep domain expertise: AI can synthesize information for outlines, but it cannot replace years of hands-on experience. We lean on AI for structure and quality control, not for expert takes.
Perfect accuracy on the first try: Even with claims ledgers and QA passes, errors slip through. You still need an editor.
Zero cost: API usage adds up. Expect to spend $50-$200/month on tokens depending on volume, plus the $6-$12/month for the VPS.

If you want to publish AI-generated SEO spam, this is not the workflow for you. If you want to publish thoughtful, fact-checked posts with AI handling the tedious planning and editing work, this might be exactly what you need.

Try It Yourself

You do not need to be a developer to set this up. If you can SSH into a server and follow a tutorial, you can build this workflow. Start small:

Spin up a $6 droplet on DigitalOcean
Install OpenClaw
Connect Telegram
Define one agent and run a test draft

See if it fits your process. Tweak the prompts. Add QA checks. Build your own agent personas. The system is flexible enough to adapt to almost any editorial workflow.

If you want to dig deeper, the full OpenClaw documentation is at docs.openclaw.ai. The GitHub repo has examples and starter configs.

This is how we ship content. No secrets, no hype. Just a reproducible system you can steal.

FAQ

Do you write everything with AI?
No. AI generates the outlines and does quality control. The actual writing is done by our authors. Every piece gets editorial review before it goes live.

How do you prevent AI from making things up?
Every claim in a draft must come from the claims ledger, which includes sources. The QA agent verifies that sources actually support the claims. If something cannot be cited, it gets cut.

What does this cost per month?
VPS: $6-$12/month. API tokens: $50-$200/month depending on volume. Total: roughly $60-$210/month for a steady publishing cadence.

Can I use this for commercial content?
Yes. We do. Just make sure your QA process is solid and you are comfortable with the legal and ethical implications of AI-assisted content in your industry.

How long did it take to set up?
Initial VPS setup and OpenClaw install: about 30 minutes. Defining agent personas and tuning prompts: a few hours of iteration. Getting the full editorial workflow dialed in: a couple of weeks of trial and error.

Want to build your own version? Start with a $6 DigitalOcean droplet and see what ships. If you run into trouble or have questions, reach out. We are figuring this out as we go, and we share what we learn.