How I Do Content Engineering with Claude Code

This process is geared specifically towards informational SEO content. I only use this process on topics that I understand well, so that I can review each article to validate its claims, correct misinformation, and make sure I feel happy putting it out into the world.

I also focus primarily on topics that Ahrefs has already covered (in some capacity), allowing us to use hundreds of existing, high-quality articles as a reference point for new content.

I could use this process to scale the Ahrefs blog to tens of thousands of articles. I will not. It would not be in the interests of Ahrefs or our customers.

Instead, I am using this workflow to help us maintain an evergreen library of useful content on a handful of core topics. My goal is to remove drudgery and focus human grey matter on the parts of marketing that benefit most from it.

At the heart of this process are ~23 skill files that correspond to different parts of the Ahrefs editorial process, from conducting keyword research to topic gap analysis to structural outlining:

Each skill file includes a Markdown-formatted explanation of how Claude (or any LLM) should conduct each process, best-practice examples to emulate, and formatting instructions for the expected output.

Many of these skills are adapted from our existing, human-written process documentation. Others are written from scratch, and some are generated and edited entirely by AI.

Every skill can be used in isolation, but I also created a main skill (blog-pipeline) that instructs the LLM to trigger each of these skills in a particular order, working sequentially through every process to take a keyword idea through to (nearly) finished article:

One risk of agentic content creation: if you get an article at the end of a ten minute run, and it’s bad, it’s hard to diagnose precisely where and why the process went wrong.

For that reason, every step of this process produces its own output file. For example, when the outline is generated, the outline is handed over to the next stage of the process, but also saved as a markdown file in the outlines folder.

I can review every single stage of the process, tweak that particular output (and the corresponding skill file), and restart from the last stage that meets my quality criteria.

As LLM model capabilities get better and better, I’m often surprised at how good frontier models are at very specific tasks, even without any concrete direction or examples provided. Sometimes, convoluted skill files are actually inferior to giving the model a single-sentence prompt and getting out of its way.

A big part of our previous AI content process was front-loading human input. My thesis is that small amounts of expert direction provided at the start of the content creation process are vastly more effective than lots of human editing at the end.

I wanted to allow for this direction without requiring a full-blown content brief to be created, so I added a context parameter to the blog-pipeline skill that allows the user to provide context to guide the content creation process.

If you want to generate an article about “content gap analysis”, you can add high-level direction like this:

“Take a ‘steal your competitors’ best content’ angle, feature Keywords Explorer’s Content Gap tool heavily, and include a section on finding quick-win keywords your competitors rank for but you don’t.”

I use this to mention specific sub-topics to cover, overarching angles and sentiments to shape the article, and specific product features to mention. This context is saved to a dedicated file and used as a reference in the drafting skill.

I read every word of every article that makes its way onto the Ahrefs blog. Reading Markdown files in VS Code isn’t my idea of a good time, so I use a simple skill that turns each generated article into a Ahrefs-style HTML preview that opens automatically in Chrome.

I’m also experimenting with interactive previews that allow me to accept or decline updates to existing content, and leave in-line comments for Claude to action. At this point, we’re straying into fully-fledged application territory, so stay tuned for the v3 of my content process.

When I’m happy with the article draft and ready to upload and add images, I trigger a final skill that formats the article with all the correct tables and shortcodes we need.

This content process is built to my personal specifications. It references my favourite articles to shape each articles tone and style. It prioritises my favorite data sources. And most importantly, it’s built to mirror my writing process, approaching content creation in a way that matches how my brain operates.

But there is no “best” way to create content. Even on the Ahrefs blog team, there are a wealth of different styles, preferences and workflows that shapes how content gets made.

I want our AI content processes to reflect all our idiosyncrasies, so I encouraged the team to fork this repo and use Claude Code to modify it to their unique specifications, adding and removing steps of the process, changing the data sources and reports it uses, and learning from the style and voice of their own best articles.

The goal is for everyone on the team to have their own personalised content copilot, able to work to their specifications and play to their strengths.

If you follow the Ahrefs blog, I’d wager that you haven’t noticed any major changes, despite the fact that we’ve been using generative AI to help with more and more of our work.

That’s because we’re not using AI to “scale our output” and publish thousands of articles, and we’re not making trade-offs by substituting quality for speed. Instead, we’re using it to automate the most formulaic parts of work, and only in those situations where AI can do the job as well as, or even better than, a skilled human could.

AI is helping us to fill content gaps and update our evergreen library of search content. In the meantime, the Ahrefs blog team can use our energy in other ways: conducting research and writing thought leadership, hosting webinars and giving in-person talks, and building systems to further automate the most tedious parts of our lives.

For all the fear-mongering around AI and content creation, this is a lot of fun.