A pipeline that takes one keyword and returns a finished WordPress draft — researched, structured, illustrated, with schema and internal links. Every article linked further down this page came out of it, and every number on this page was read today from the live site's API and from the job records the pipeline writes.
Framing first, because it matters: this is my own site, not a client commission. There are no client logos or testimonials on this site because I do not have any to show yet. What I can show is a system that has been running on its own schedule for months and the output it produced, which is more useful to a buyer than a quote attributed to someone you cannot call.
A site needs articles faster than one person can research and write them, and slower than a content mill produces them, because the second kind gets the site penalised. The task was not "generate text". It was "produce something I would be willing to put my name on, at a pace I could not sustain by hand".
Research, outline, draft, edit, source images, write meta and schema, then publish — the better part of a working day per article when done properly, and the part of the day that gets skipped first when client work arrives.
A complete draft in roughly fifteen minutes, at a length and structure set by the keyword rather than by a template: release notes come out short, comparison guides come out past three thousand words.
Nothing publishes itself. The pipeline delivers a draft to WordPress and stops; the publish decision, the fact-check and the final edit are mine. That is a design choice, not a missing feature.
The interesting part of a pipeline is not the happy path. It is which stage is allowed to say no, and what happens when it does.
Picks from a store of 22,382 keywords by buyer intent and difficulty rather than by whatever sounds interesting, and skips anything already covered. Refuses to run if nothing in the queue clears the bar — a quiet day is better than a filler article.
Queries a knowledge base — vector search plus full-text — for anything already written about the topic, so the new article does not repeat the last one in different words. This is the stage most content automation skips, and it is why those systems produce forty variations of the same post.
Collects what it can verify, builds an outline with a table of contents and internal links, and records the sources it used so a claim can be traced back afterwards. Refuses to proceed on a topic where it found nothing authoritative.
Writes the article to the outline, in a fixed voice, with meta description, headings and FAQ block. Length follows the topic, not a word target — padding to hit a number is what makes generated content obvious.
Rules plus a model pass: word count floor, duplication ratio, banned phrases, structure checks, and a readability judgement. If it fails, the article is rejected and the reason is logged. The gate exists to throw drafts away — a gate that always passes is decoration.
Header and inline images generated or selected, sized and compressed for the web, with alt text written from the section they sit in. If image generation fails, the article still goes out — a missing image is not a reason to lose a day.
Pushed to WordPress over REST as a draft with schema, categories and tags attached. Every stage writes its own record, so when something looks wrong on the site I can trace which stage produced it instead of guessing.
A pipeline nobody watches fails in silence. These are the failure modes I have actually hit, and what the system does about each one.
Retried with a different model or after a wait, not with the same request repeated. If every fallback fails, the run stops and says which stage and why, rather than delivering a half-built article.
Rejected at the gate, with the failing check named in the log. Some topics genuinely do not produce a publishable article from a machine, and the honest response is to drop that keyword rather than to lower the gate.
WordPress REST does not always accept what a plugin expects, and a silent 200 with an empty body is a real failure mode. The delivery stage confirms the post exists and reads it back before recording success.
Separate weekly job re-checks the third-party sources the articles cite, diffs them, and flags the pages that quote a figure which has since changed. Content that is wrong is worse than content that is missing.
Image generation on a local GPU competes for the same memory as everything else on the machine. A guard job stops idle model containers so the pipeline is not the job that fails at three in the morning.
The real failure mode of every automation. Failures are pushed to me as a message with the diagnosis attached, not written to a file I would have to remember to open.
Read them as writing, then remember a pipeline produced the first draft. Lengths range from 879 words for a product release note to 3,803 for a pricing review; nine of the sixteen run past 2,000 words. Median images per article: two.
Plans, limits and what the credits actually buy, with the free tier tested rather than described.
command-code-pricing-reviewIndexing custom fields and custom post types in WordPress search, including where the default query gives up.
wordpress-search-custom-fieldsFree chatbot plugins compared on what they let you control, not on a feature checklist.
best-free-chatbot-for-wordpressWhy a bring-your-own-key chatbot changes the cost model, with the arithmetic shown.
self-hosted-chatbot-byokWiring a site chatbot into n8n so conversations become leads instead of transcripts nobody reads.
chatbot-with-n8nThe short format, included deliberately: not every post should be three thousand words.
yakwp-2-4-0-releaseNo interviews, no original research, no first-hand testing beyond what a machine can reach. Articles that need a source on the phone are not this system's job.
It produces a technically sound article. Whether Google rewards it is a separate question with a months-long feedback loop, and anyone quoting you a ranking guarantee is selling you something they cannot deliver.
By design. A human reads the draft, checks the claims and presses publish. Automating that step is how sites end up with a factual error indexed for two years.
The build is two to three weeks. The first step is a paid diagnosis, because I would rather tell you which parts of your content process can be automated than discover it halfway through a build.
Site access, the topics you actually want to rank for, three pieces of writing you are happy with so the voice can be matched, and a decision on whether drafts publish themselves or wait for you.
Your CMS, whether the model runs on your hardware or a rented one, your language, and how much fact-checking the topic needs. Those four decisions change the build more than anything else.
Nothing per article beyond compute, because the models run on hardware you control rather than per-token through an API. That is the difference between a pipeline that survives a busy month and one that gets switched off.
Send the site and the topics you are trying to own. You will get an honest answer about what can be automated, what cannot, and what it takes.