Right here’s how I constructed a high-quality content material automation system for the Ahrefs weblog utilizing Claude Code and 23 ability information.
Again in August 2025, I shared the AI content material course of I had developed for the Ahrefs weblog. It used ChatGPT initiatives and customized GPTs to hurry up sure kinds of content material creation from a number of days to a few hours, however nonetheless required tons of guide intervention.
Now, barely eight months later, I’m sharing our new course of. I exploit Claude Code and 23 customized ability information, chained collectively, to generate publish-ready article drafts in six to 12 minutes. We now have printed round 15 articles with this new course of, and up to date some 30 or so extra.
I’ve been utilizing AI to assist create content material advertising and marketing since 2020. It has been helpful in effortful, piecemeal methods. However as we speak it’s ok to automate necessary elements of content material advertising and marketing with no loss in high quality (and even a major acquire in some areas, like analysis). Or as I put in a latest article: AI content material wasn’t ok. Now it is.
Because of this, I urged a reasonably daring course in our firm Slack, again in February:
Right here’s our AI present content material course of.
Watch this course of on YouTube
Take a look at this episode of the Ahrefs podcast to look at me demo our content material automation system to Ahrefs’ CMO, Tim Soulo.
Earlier than we get to the great things, I as soon as once more need to direct your consideration to some necessary caveats:
Expertise issues
AI content material isn’t, by default, good. This course of works nicely as a result of it mirrors our current human editorial course of, constructed from many years of collective content material advertising and marketing expertise. Or as somebody in a LinkedIn remark put it, very articulately:
“Ryan’s SKILL information are good as a result of Ryan already knew what to place in them. Most individuals utilizing blank-slate instruments don’t have 13 years of editorial expertise to construct from. The hole isn’t simply within the instrument. It’s within the particular person behind it too.”
Subject choice nonetheless issues
This course of is geared particularly in direction of informational search engine optimisation content material. I solely use this course of on matters that I perceive nicely, in order that I can evaluation every article to validate its claims, appropriate misinformation, and ensure I really feel joyful placing it out into the world.
I additionally focus totally on matters that Ahrefs has already coated (in some capability), permitting us to make use of lots of of current, high-quality articles as a reference level for brand spanking new content material.
We now have no plans to “scale content material” with AI
I might use this course of to scale the Ahrefs weblog to tens of 1000’s of articles. I can’t. It might not be within the pursuits of Ahrefs or our clients.
As a substitute, I’m utilizing this workflow to assist us keep an evergreen library of helpful content material on a handful of core matters. My purpose is to take away drudgery and focus human gray matter on the elements of promoting that profit most from it.
On the coronary heart of this course of are ~23 ability information that correspond to completely different elements of the Ahrefs editorial course of, from conducting key phrase analysis to matter hole evaluation to structural outlining:
Every ability file features a Markdown-formatted rationalization of how Claude (or any LLM) ought to conduct every course of, best-practice examples to emulate, and formatting directions for the anticipated output.
Many of those abilities are tailored from our current, human-written course of documentation. Others are written from scratch, and a few are generated and edited solely by AI.
Each ability can be utilized in isolation, however I additionally created a principal ability (blog-pipeline) that instructs the LLM to set off every of those abilities in a selected order, working sequentially by means of each course of to take a key phrase concept by means of to (practically) completed article:
Theoretically, this course of will be totally automated. Utilizing the ability information I created, Claude can set off a every day content material hole evaluation utilizing the Ahrefs MCP, evaluation and prioritize the most effective key phrases to focus on, and kick off all the blog-pipeline workflow, notifying me when new article drafts are prepared for evaluation.
One threat of agentic content material creation: for those who get an article on the finish of a ten minute run, and it’s dangerous, it’s laborious to diagnose exactly the place and why the method went incorrect.
For that purpose, each step of this course of produces its personal output file. For instance, when the define is generated, the define is handed over to the following stage of the method, but additionally saved as a markdown file within the outlines folder.
I can evaluation each single stage of the method, tweak that individual output (and the corresponding ability file), and restart from the final stage that meets my high quality standards.
As LLM mannequin capabilities get higher and higher, I’m usually shocked at how good frontier fashions are at very particular duties, even with none concrete course or examples supplied. Typically, convoluted ability information are literally inferior to giving the mannequin a single-sentence immediate and getting out of its manner.
We’ve been utilizing Anthropic’s skill-creator ability to check and enhance our workflow. The ability checks every stage of the method, producing outlines, analysis primers and drafts each with and with out the steerage supplied in our customized ability information.
The LLM critiques the outputs and makes strategies for learn how to enhance the ability file for extra constant outcomes.
It’s simple for ability information to turn into lengthy and bloated, and in doing so, make it much less seemingly that their steerage will probably be accurately utilized by the LLM. This course of permits me to repeatedly shorten abilities to their only essence, and take away abilities that don’t have any actual bearing on my desired output.
This course of is barely potential as a result of Claude has entry to the Ahrefs MCP.
As a substitute of hallucinating faux search engine optimisation information, Claude can pull key phrase metrics, dad or mum matter, and long-tail key phrase variations for each article, straight from Ahrefs.
It makes use of the questions report back to floor generally requested questions and teams them into themes, and it retrieves the SERP overview to know the dominant search intent and what sort of content material is rating.
In addition to nice search engine optimisation information, my ability information additionally embrace particular directions to make use of different necessary information sources, like:
- Competitor information: key matters, headers and content material gaps are extracted from top-ranking articles on the identical key phrase.
- Deep analysis: trusted information and analysis sources are reviewed for latest info on the goal key phrase.
- Product options: the LLM has entry to an summary of each Ahrefs product and have, saved in a Markdown doc, together with their most necessary use instances.
By default, LLMs are very convincing bloviators: they’ll generate content material that sounds coherent, with out containing any concrete information or substance. Mandating particular information sources to make use of is essential to getting nice outcomes.
Get Ahrefs information in your AI instruments
The Ahrefs API and MCP at the moment are accessible in all paid plans from Lite and above, so it’s simpler than ever to get world-class Ahrefs search engine optimisation and AEO information in your dashboard, utility or content material workflow.
Want some inspiration to get began? Learn this: 15 Ahrefs MCP Use Instances for SEOs & Digital Entrepreneurs
A giant a part of our earlier AI content material course of was front-loading human enter. My thesis is that small quantities of knowledgeable course supplied on the begin of the content material creation course of are vastly simpler than a lot of human modifying on the finish.
I needed to permit for this course with out requiring a full-blown content material transient to be created, so I added a context parameter to the blog-pipeline ability that enables the consumer to offer context to information the content material creation course of.
If you wish to generate an article about “content material hole evaluation”, you may add high-level course like this:
“Take a ‘steal your opponents’ finest content material’ angle, function Key phrases Explorer’s Content material Hole instrument closely, and embrace a piece on discovering quick-win key phrases your opponents rank for however you don’t.”
I exploit this to say particular sub-topics to cowl, overarching angles and sentiments to form the article, and particular product options to say. This context is saved to a devoted file and used as a reference within the drafting ability.
I learn each phrase of each article that makes its manner onto the Ahrefs weblog. Studying Markdown information in VS Code isn’t my concept of a superb time, so I exploit a easy ability that turns every generated article right into a Ahrefs-style HTML preview that opens mechanically in Chrome.
Sidenote.
As you may see within the picture above, weblog submit photographs aren’t a solved drawback, but—I’m at present experimenting with abilities to set off a headless browser that may navigate to particular Ahrefs stories, take screenshots, annotate them, and insert them into the article draft.
I’m additionally experimenting with interactive previews that enable me to just accept or decline updates to current content material, and go away in-line feedback for Claude to motion. At this level, we’re straying into fully-fledged utility territory, so keep tuned for the v3 of my content material course of.
After I’m pleased with the article draft and able to add and add photographs, I set off a last ability that codecs the article with all the right tables and shortcodes we want.
This content material course of is constructed to my private specs. It references my favorite articles to form every articles tone and magnificence. It prioritises my favourite information sources. And most significantly, it’s constructed to reflect my writing course of, approaching content material creation in a manner that matches how my mind operates.
However there isn’t any “finest” strategy to create content material. Even on the Ahrefs weblog staff, there are a wealth of various kinds, preferences and workflows that shapes how content material will get made.
I need our AI content material processes to mirror all our idiosyncrasies, so I inspired the staff to fork this repo and use Claude Code to change it to their distinctive specs, including and eradicating steps of the method, altering the info sources and stories it makes use of, and studying from the model and voice of their very own finest articles.
The purpose is for everybody on the staff to have their very own personalised content material copilot, in a position to work to their specs and play to their strengths.
The longer term
For those who observe the Ahrefs weblog, I’d wager that you just haven’t seen any main modifications, even if we’ve been utilizing generative AI to assist with increasingly of our work.
That’s as a result of we’re not utilizing AI to “scale our output” and publish 1000’s of articles, and we’re not making trade-offs by substituting high quality for velocity. As a substitute, we’re utilizing it to automate essentially the most formulaic elements of labor, and solely in these conditions the place AI can do the job in addition to, and even higher than, a talented human might.
AI helps us to fill content material gaps and replace our evergreen library of search content material. Within the meantime, the Ahrefs weblog staff can use our vitality in different methods: conducting analysis and writing thought management, internet hosting webinars and giving in-person talks, and constructing programs to additional automate essentially the most tedious elements of our lives.
For all of the fear-mongering round AI and content material creation, this can be a lot of enjoyable.
P.S. I’m already constructing model three of this content material system, and it’s one thing you should use too. Be a part of this waitlist to get early entry.
