Three months ago I was writing every article by hand. Researching topics, drafting, formatting for WordPress, generating images, scheduling social posts. Each article ate 3 to 4 hours.
Today the same pipeline runs mostly automated. Six articles per session. About 100 minutes total, most of which I spend reviewing rather than writing.
Here is the full architecture.
The Core Problem This Solves
Most content advice says "publish consistently." Nobody tells you how to do that when you are a solo operator with a day job or other projects eating your time.
The answer is not to write faster. It is to build a system where writing is only one small part of the pipeline, and most of the other parts run without you.
The Stack
Claude (Sonnet-class for cost, Opus for synthesis) handles all writing. Every skill in the system is a Markdown file with structured instructions. They chain together like functions: keyword picker feeds article writer, article writer feeds auditor, auditor feeds publisher. Each skill gets exactly the context it needs, nothing more.
Google Search Console integration is the highest-leverage piece. The GSC API returns every keyword where the site ranks position 5 to 20. These keywords already have traction. The site is indexed for them. The top result is reachable. Writing a targeted article around these terms has a much higher expected return than guessing what to cover.
The picker also screens against the live sitemap to avoid cannibalization. If a published article already targets a keyword closely, it gets filtered out.
Perplexity API feeds the news slot. Every session it surfaces today's breaking AI news ranked by novelty and search volume. One news article gets written per session.
fal.ai FLUX.1 generates images. Three images per article, from prompts the article writer produces. The constraint: no text in image prompts. FLUX.1 cannot render legible text reliably. Use visual metaphors instead. A prompt like "a solitary robot at a desk under fluorescent office lighting, cinematic, dramatic shadows" works. "Robot writing AI SEO articles text overlay" does not.
WordPress REST API handles draft publishing. The pipeline creates the post with all metadata: title, body, featured image, categories, tags, excerpt, canonical URL. A human (me) then spot-checks and clicks publish. That step stays manual deliberately. Automation handles the tedious parts. Judgment stays with me.
Twitter API (OAuth1) handles distribution. After an article publishes, a Python script converts it into a 3 to 4 tweet thread automatically and posts it.
The Workflow Step by Step
1. Keyword picker queries GSC for position 5 to 20 gaps. Returns up to 4 picks: one Review, one Comparison, one Tutorial, one AI Companion update. Screens against sitemap.
2. News sourcing hits Perplexity for today's breaking AI stories. Picks the one with the most original angle.
3. Article writer runs on all 6 articles in sequence. For each: pulls competitor context, drafts full 1,500-word article with internal links, affiliate citations, a structured summary box, and at least one markdown table where appropriate.
4. Auditor catches mechanical issues before they reach WordPress: banned phrases, missing structural elements, title length out of range. Fixes what it can in-place.
5. Image generator runs on each article's FLUX.1 prompts.
6. Publisher uploads all 6 to WordPress as drafts via REST API.
7. Content distributor runs on all recently published articles. Twitter thread for each. Hacker News for the two best fits. Batches the rest.
What the System Cannot Do
Building links. The pipeline publishes and distributes but cannot make other sites link to you. That is still manual and community-based.
Genuine personal authority. Google increasingly values real author expertise. The system can write well at scale but cannot inject lived experience it does not have. You need to add that in review before publishing.
Quality judgment calls. The auditor catches mechanical problems. It cannot tell you whether an article is genuinely better than the top 3 results for that keyword. That call is yours.
The Numbers
The site this runs on hit a rough patch in late 2024. A Google core update cut traffic by roughly 60 percent. Running this pipeline daily over the past few months has rebuilt topical authority slowly.
Position 5 to 20 keyword count went from about 180 to over 400. Several articles have climbed into position 3 to 8 for their targets. It is a slow process, but the trajectory is right.
What I Packaged From This
I documented the full system: the agent architecture, the prompt patterns that actually work, the keyword selection logic, the WordPress and Twitter integrations, the image generation setup, and the distribution flow.
It is a practical guide for solopreneurs who want to run a serious content operation without a team. Not theory. The exact system I use daily.
If that is useful to you: The Solopreneur's AI Content Engine at $27.
Questions on any part of the stack? Drop them in the comments. Happy to go deeper on the GSC integration or the agent chaining approach specifically, since those are the two pieces that took the most iteration to get right.
Top comments (0)