Full Automated AI Batch Writing Workflow

After running a content site for a few months, I hit a real problem: writing articles by hand is just too slow. A single article, from topic selection to final draft, takes about two hours of careful work. Five articles a day means ten hours gone — nothing else gets done. What's more anxiety-inducing is that content sites need a critical mass of articles to trigger compound effects. Under 50 articles, you get basically no traffic. At 100, things start moving. At 200+, you can consistently earn organic search traffic. At manual speed, that's 60 articles a month — three months to reach 200. That's way too slow in the competitive SEO landscape.

So I started exploring: can I use AI to automate article writing? After months of trial and error and optimization, I built a complete AI batch writing workflow — from topic selection to publishing, fully automated. Efficiency improved at least 10x. Where a single article used to take 2 hours manually, this workflow produces 10 articles in 30 minutes. And the quality matches hand-written work. In this article, I'll break down every step of my workflow: tool selection, prompt design, SEO optimization strategies, image generation, and how to achieve full-chain automation from git push to auto-deployment.

Why This Topic Matters

The first step in the entire workflow is keyword research.

Don't pick keywords by gut feeling — use data.

My primary tools are Google Search Console (GSC) and a free tool called Keyword Tool.

GSC's "Search Results" report shows which keywords are driving impressions and clicks for your site.

Look for keywords with low click-through rates but high impressions — this means your content ranks but isn't compelling enough.

Optimizing those can yield improvements.

I also export all keyword data from GSC's "Queries" report to analyze what users are actually searching for.

For example, I noticed that "sports suit for wedding" had an 8% CTR — a very clear search intent.

Users are looking for sports suits to wear to weddings.

So I could batch-produce a series of "What to wear for [occasion]" articles.

Once keywords are selected, step two is building standardized content templates.

I wrote a complete set of Prompt templates to solve the problem of inconsistent AI content quality.

The core principle is: the more detailed your instructions to the AI, the better the output.

My standard template includes: target keyword, article type (product review / buying guide / scenario recommendation / comparison analysis), word count requirement (3000+ words), target reader profile, tone and style (friendly, professional, data-backed), paragraph structure requirements (150 to 300 words per paragraph, no repetition, data and case studies included), and SEO requirements (naturally integrate keywords, include H2/H3 headings, write Meta descriptions).

Here's an actual prompt I use: "Write an article about 'What occasions are sports suits suitable for.

' Target readers are working men aged 25 to 40 who want to buy a sports suit but aren't sure when to wear one.

The article should include: what daily occasions are suitable for sports suits, whether they can be worn for formal occasions, what colors suit what occasions, and price range recommendations.

Each paragraph 200 to 300 words, total article 3000+ words.

Naturally integrate keywords 'sports suit occasion,' 'sports suit formal occasion,' 'sports suit daily outfit.

' Tone should be friendly and professional, like an experienced styling consultant sharing advice.

Don't use numbered lists — use natural paragraph breaks.

End with a summary." Every article generated this way is complete and substantial — not the thin AI garbage you see everywhere.

Step 1: Find Your Positioning

Step three is batch generation. I typically prepare 20 to 30 keywords at a time in a CSV table, one keyword per row. Then I use a script to call the ChatGPT API in batches to generate articles. A key insight: don't generate too many at once. Best to do batches of 5, do a manual review after each batch before continuing. Two reasons: first, API rate limits; second, batch-generated content quality varies and needs human quality control. After each article is generated, I do a quick review — check for factual errors, reasonability of arguments, structural completeness. This takes about 3 to 5 minutes per article, way faster than writing from scratch.

Step four is SEO optimization.

After the AI-generated draft is done, some SEO-level optimization is needed.

First is internal linking.

Each article should link to at least 2 to 3 related articles within the site.

For example, when writing "What occasions are sports suits suitable for," I link to already-published "Sports Suit Fabric Buying Guide" and "Sports Suit vs.

Regular Suit Differences.

" This improves user experience and helps search engine crawlers go deeper into your site.

Second is Meta information optimization — title tags, Meta descriptions, URL slugs.

Titles should be within 60 characters, Meta descriptions within 160 characters, both containing core keywords.

URLs use English slugs, kept short.

For images, I use DALL-E exclusively. A high-quality product scene image costs under $0.05 each. The key: don't use those obvious AI-generated abstract images — generate photorealistic product photos. My DALL-E prompt example: "professional product photo of a navy blue tailored sport coat worn by a man in a coffee shop, natural lighting, realistic texture, high resolution, commercial photography style, white background inset." English prompts typically produce more natural results. Each article gets 2 to 3 images, differentiated by scene. Visual cost per article is $0.10 to $0.20 — hundreds of times cheaper than stock photo libraries.

Step 2: Build the System

After generating images, two more processing steps. First, compression — use TinyPNG or Squoosh to batch compress images to under 200KB, keeping page load times fast. Second, add Alt tags — every image needs descriptive Alt text, which is both an SEO requirement and an accessibility requirement. Alt text should naturally contain keywords without stuffing. Example: "Gray sports suit paired with dark blue jeans in a coffee shop."

Step five is pushing content to the GitHub repository to trigger build and deploy. This is the critical chain link in the entire automation process. I have a GitHub repository with MDX files organized by topic. Each article is one file, named slug.md. File content includes frontmatter (title, slug, description, tags, date, and other metadata) plus the body in Markdown. After writing the file, I use Git to commit and push. Vercel has GitHub integration configured — every push to the main branch automatically triggers build and deployment.

GitHub Actions does even more here. I wrote a workflow file that, after each push, automatically: checks all markdown files for dead links (internal and external); auto-optimizes images by compressing large ones to WebP format; runs ESLint for code quality checks; and only if all checks pass, triggers Vercel deployment. I just need to write content and push — everything else is automated.

Step 3: Content Output

Let's compare efficiency data.

Traditional manual writing workflow: topic selection 30 min, research 1 hour, writing 2 hours, editing and proofreading 30 min, images 30 min, upload and publish 30 min.

Total 4 to 5 hours per article — at most 2 to 3 articles per day.

My AI automation workflow: batch keyword research — 20 min for 10 keywords; generate 10 articles with template — about 15 min API calls; manual review of 10 articles — about 30 min; SEO optimization and images — about 20 min; git push to deployment — about 2 min.

Total under 2 hours for 10 articles.

Efficiency improved 10 to 15 times.

And this process is 100% repeatable and scalable — 10 articles today, 10 articles tomorrow, no decline from physical exhaustion.

Some might question: is AI-generated content quality good enough? Will Google penalize AI content? I researched this thoroughly, including Google's official guidelines. Google's position is clear: they don't penalize AI content per se — they penalize low-quality content. Whether written by humans or AI, as long as the content is valuable to users, original, and in-depth, Google will rank it well. The key is avoiding keyword-stuffed junk with no real value. So my AI writing process heavily emphasizes content quality: every article must include real data, specific case studies, actionable advice, and insights based on my personal experience. AI is an assistant, not an excuse to stop thinking.

Of course, AI-generated content isn't perfect. I've hit several pitfalls. First, AI sometimes fabricates data. It might say "According to a study, 85% of users choose this way" — but that study may not exist. So I check data sources in every generated article and remove unverifiable claims. Second, AI content lacks depth of personal perspective. It might cover a topic comprehensively but miss personal experience and judgment. So I require at least one or two paragraphs based on my actual experience, which I add during review. Third, AI sometimes repeats itself or has logical leaps — manual fine-tuning is needed.

Step 4: Traffic Acquisition

About Prompt iteration: my advice to beginners is don't expect to write the perfect Prompt on the first try. Write a first draft, generate an article, see how it looks, then adjust. One effective technique: add "forbidden" instructions, like "Don't use generic opening lines''" or "Don't start every paragraph with 'first,' 'second,' 'last.'" Another technique is to provide examples — include an article you're proud of as a reference and say "Write an article about [topic] in this style." The output quality noticeably improves.

Another thing to watch is originality.

The most easily overlooked problem with batch-generated content is "semantic repetition" — different articles with different keywords but nearly identical structure and viewpoints.

Google is very good at detecting patterns like this, so ensure each article is independently high-quality.

My approach: give every article a unique angle.

For sports suits, article one covers fabric, article two covers wearing occasions, article three covers price ranges, article four covers brand comparisons, article five covers care and maintenance.

Each article has a completely different entry point and information structure — Google can't possibly flag them as duplicate content.

If you don't have programming skills, here's a code-free alternative: use Notion to manage keywords and content, then connect Notion to GitHub via Zapier or Make so articles written in Notion automatically sync to your GitHub repo. Or even simpler, use Ghost CMS or WordPress as your content management backend, with an AI plugin generating articles directly in the backend. Less flexible than the Next.js approach, but a viable automation path for non-coders.

Practical Case Study

Finally, the ROI of this workflow.

My cost breakdown: ChatGPT API — about $20 to $30/month for 100 articles; DALL-E image generation — about $10 to $15/month; domain name — under $6/month ($45/year); server cost — zero.

Total monthly cost: $30 to $40.

The output: 255 articles accumulated in three months; daily UV grew from 0 to 500+; total site impressions went from zero to tens of thousands per month.

Based on conservative estimates from AdSense and CPS affiliates, monthly revenue is expected to be between 500 and 2000 RMB — covering all costs within six months.

The core value of this workflow isn't single-time efficiency gains — it's enabling consistent, stable, high-quality content output.

That's what determines whether a content site can scale.

Summary of the workflow's core points: keyword research with GSC + Keyword Tool to find long-tail keywords; content generation with ChatGPT + standardized Prompt templates; images with DALL-E for product scene photos; SEO optimization with internal links and Meta information; automated deployment via git push triggering GitHub Actions + Vercel. Once you master this entire flow, producing 10 to 15 high-quality articles per day is easy. With content quantity and quality secured, traffic and revenue follow naturally.

Long-term Strategy