AI Blog Writer — Research, Draft, Check, Publish
A WordPress plugin that researches a subject, drafts a post, criticises its own draft and checks the result before it reaches you. You bring your own API key, so there are no credits and no monthly quota.
Most AI writing plugins do one pass and hand you the average
Ask a model for an article and it returns the midpoint of everything already published on that subject. That is precisely the thing you are trying to outrank.
The gap is not the model. It is that nothing checks the output, nothing researches first, and nothing refuses to publish something weak. This plugin is four passes rather than one, and the last two exist to throw work away.
Four passes
-
Research
Gathers source material before writing a word. Optional, off by default, and it uses whichever search source you connect.
-
Write
Builds an outline first, then writes section by section rather than producing a wall in one go.
-
Critique and revise
Reads its own draft for vague, repetitive and padded writing, then rewrites what it flagged. If nothing is flagged the pass is skipped rather than churning.
-
Verify
Checks that links actually resolve and scores the draft against 25 checks. Anything below your threshold is held for review instead of published.
Sounding like you rather than like an AI
Style matching from a real article
Paste something you wrote. It measures the shape: length, section count, sentence length, link density. Those become writing rules. No wording is copied.
A blocklist of AI tells
Common AI phrasing is blocked by default, and you can add your own. Everybody has a word that makes them wince.
Your anecdotes
You can supply real examples for it to use. This is the difference between a page worth reading and a competent summary.
Topics to avoid
Tell it what not to write about, and it refuses topics too close to posts you already have, which is how sites end up competing with themselves.
The limits, stated plainly
It will not find your topics
You decide what to write about. Keyword research is a judgement call about your business, not a generation task.
It will not invent evidence
Figures and claims come from you. A model asked for a statistic will produce a plausible one, which is worse than none.
It defaults to drafts
Nothing publishes itself unless you turn that on, volume limits start conservative, and there is a monthly token cap.
Your key, your choice of model
Cloud providers
OpenAI, Anthropic, Google Gemini, Groq, Mistral, DeepSeek, OpenRouter, xAI, Moonshot, Together, Fireworks, Cerebras and Hugging Face, each with your own key.
Or nothing leaves your server
Ollama, LM Studio, Jan, llama.cpp or any OpenAI-compatible endpoint. For anybody who cannot send client material to a third party, this is the point.
Or your existing subscription
Claude or ChatGPT over the Model Context Protocol, so you use the plan you already pay for.
SEO fields filled in
Writes to Yoast, Rank Math or SEOPress when one is active, and outputs its own description and sharing tags when none is.
What the 25 checks actually test
The opening answers the question
Within the first two sentences. This is the part that gets quoted in a featured snippet or an AI answer, and the part most writing gets wrong by warming up first.
Figures have a source beside them
Any section stating a number must link to where the number came from, in that section rather than in a list at the bottom.
The subject is where it needs to be
In the title, in a heading, in the URL slug and in the first hundred words. Not stuffed everywhere else.
It is not just restating the research
The draft is compared against its own sources. Summarising them back costs points, because a summary of what already ranks is not a reason to publish.
Your figures survived
Numbers you supplied have to appear on the finished page. Rewrites lose things, and this catches it.
A check that cannot be assessed is skipped rather than failed, so the score is points earned out of points that could be earned. Every failed check is handed back into the rewrite as an instruction, which is the difference between scoring a draft and improving one.
Two things the checks do not do
The figure check does not open the links
It confirms a source is cited, not that the source says what the sentence claims. That part is still yours.
The originality check is not plagiarism detection
It compares the draft against the sources it was given, and nothing else. The documentation says so in those words rather than implying more.
How the style learning works
It measures, it does not copy
Sentence length, paragraph length, how often you use an em dash or a contraction, whether you write in first or second person. From an article you admire it takes only the structure. No wording is copied and none is sent to a model.
It will not learn from itself
Posts the plugin wrote are excluded from that learning on purpose, so the voice does not drift further from yours with every post.
The field that matters
The experience box. The documentation calls it the only part of a post a model cannot produce, which is exactly right, and it is why the plugin asks you for facts rather than inventing them.
Nothing saves until you confirm
It drafts a voice description from your own titles and shows it to you first.
Images, markup and old posts
Images land in your Media library
Downloaded rather than hotlinked, so they do not vanish when somebody else’s site changes. Alt text comes from the heading each image illustrates.
It stands down for your SEO plugin
With one of ten named SEO plugins active it fills their title and description fields instead of writing its own head tags. It also warns you if your theme already outputs Article markup, because two sets are worth less than one.
It links older posts to new ones
Publishing adds links from existing posts, which is the half of internal linking everybody forgets.
Posts are ordinary posts
Written straight into WordPress, editable by hand, and they stay if you remove the plugin. Deleting it keeps your settings and posts unless you tick the box that erases them.
Things worth knowing
Research sources start switched off
All of them. Turning one on is how you consent to it. Wikipedia and Hacker News need no key; Tavily and SerpApi are paid; a self-hosted SearXNG is free.
Scheduled posts need real cron
WordPress only runs scheduled tasks when somebody visits. On a quiet site, set a system cron job or posts arrive late.
Connected-client mode needs public HTTPS
Using your Claude or ChatGPT subscription instead of an API key will not work on a laptop install. Plain permalinks also break the sign-in step, and some hosts strip the Authorization header, which a CGIPassAuth line in .htaccess fixes.
Tokens are shown once
Connected apps get a token stored only as a fingerprint, revoked automatically if that person loses write permission. They can draft, score and publish above threshold, but cannot schedule or edit posts they did not create.
The full setup walkthrough, provider list and troubleshooting are in the AI Blog Writer documentation.
Will Google penalise AI content?
Google judges the page, not how it was written. Generic content does badly because it is generic. The critique and verify passes exist for exactly that reason, but they cannot supply the experience only you have.
Do I need to pay for credits?
No. You connect your own API key and pay that provider directly, or run a local model and pay nothing.
Does it send anything to you?
No analytics and nothing to the author. API keys are encrypted before storage and shown masked.
Can it publish on a schedule?
Yes, though WordPress only runs scheduled tasks when somebody visits the site. On a quiet site, set a real cron job.
What happens to a draft that scores badly?
It is set to pending and listed under Needs review, even if you chose to publish immediately. Each failed check is fed back into the rewrite as an instruction rather than just reported.
Is any of it locked behind a paid tier?
No. Every feature is in the free plugin.
Want the strategy as well as the drafts?
A plugin writes posts. Deciding which posts are worth writing is the other half, and the one that decides whether any of it works.