Generative AI has made producing written content dramatically faster. It has also made producing enormous amounts of generic content dramatically easier.

Those are not the same thing.

I use AI extensively in my professional work, and increasingly I use it to help maintain my own websites. But I do not want an AI system to simply generate articles and fill a website with them.

The goal is different.

I want AI to remove the mechanical work around writing while leaving the ideas, opinions and personal experience mine.

That distinction has become increasingly important as the tools have improved.

Start With the Idea, Not the Prompt

My preferred workflow does not begin by asking an AI model to "write a blog post about AI."

That produces exactly the kind of content the internet already has too much of.

I start with something I actually want to say.

It might come from a problem we encountered in a project, something I learned while testing a new technology, a discussion with a colleague, or simply an observation I think is worth developing.

AI can then help me explore the idea.

For example, I can ask:

  • What aspects of this topic am I overlooking?
  • What questions would a reader probably have?
  • Is my argument logically consistent?
  • Which parts need more explanation?
  • What would make this useful rather than merely interesting?

At this stage AI works much better as a thinking partner than as an author.

Research Is Where AI Saves Me the Most Time

Research used to be one of the slowest parts of writing technical articles.

Search for information. Open ten browser tabs. Read documentation. Compare sources. Try to remember where a particular number or statement came from.

AI-assisted research can compress much of this work.

But there is an important rule: the more specific or current the claim, the more important it is to verify the source.

I am comfortable using AI to explain a concept I already understand.

I am much less comfortable publishing a current product limitation, price, benchmark or technical specification simply because a model says it is true.

AI makes research faster. It does not eliminate the need for judgment.

Let AI Create Structure

Once I know what I want to say, AI is very useful for organization.

A rough collection of ideas can quickly become:

  • an introduction,
  • a logical sequence of sections,
  • examples,
  • conclusions,
  • and possible titles.

This solves one of the annoying problems of writing: staring at an empty page.

It is much easier to improve an imperfect structure than to create the first structure from nothing.

I rarely accept the first outline unchanged. The value is that I now have something concrete to react to.

The First Draft Can Be Collaborative

This is where workflows become more interesting.

Sometimes I write a section and ask AI to improve clarity.

Sometimes I explain what I want to say and let AI produce a first version.

Sometimes AI produces most of a section and I rewrite it heavily.

There is no single correct method.

What matters is that I remain responsible for the final text.

If a paragraph does not sound like something I would say, I change it.

If it contains an opinion I do not hold, it disappears.

If it sounds impressive but says very little, it definitely disappears.

AI is very good at producing grammatically polished sentences. That does not automatically make those sentences worth publishing.

Keeping a Personal Voice

This is probably the biggest challenge with AI-assisted writing.

Language models naturally tend toward polished, balanced, generic prose. If left alone, different authors can begin to sound strangely similar.

Personal writing needs imperfections.

It needs opinions.

It needs specific experiences.

It sometimes needs a short sentence where a perfectly optimized content generator would prefer a paragraph.

The easiest way I have found to preserve voice is to provide the substance myself.

AI can help express the idea, but the idea needs to come from somewhere real.

For this article, for example, the workflow I am describing is the workflow I actually use. That gives the text an anchor that a generic prompt cannot provide.

AI Can Handle the Boring Publishing Work

Writing is only part of publishing a blog post.

There is also:

  • formatting,
  • creating a slug,
  • choosing a category,
  • writing metadata,
  • checking links,
  • preparing an image,
  • adding alt text,
  • updating search,
  • updating the sitemap,
  • validating the site,
  • and deploying it.

This is exactly the kind of repetitive, rule-based work that an AI coding agent can handle well.

My current websites are static Next.js applications stored in GitHub and deployed automatically to Azure.

Instead of maintaining a traditional CMS, I can define a publishing skill that tells an AI coding agent exactly how a new article should be added.

The workflow becomes remarkably simple:

I provide the article or idea.

The agent creates the content file, validates the metadata, checks the image and links, runs the build and confirms that everything is ready.

After review, the change is committed to GitHub and deployed automatically.

In effect, the AI agent becomes a very lightweight CMS.

Automation Still Needs a Boundary

I do not want every step automated.

In particular, I prefer a clear boundary before publication.

The agent can create the post.

It can validate it.

It can build the website.

But publishing should still require an explicit decision.

That gives me a final opportunity to read the article as a reader rather than as someone who has been working on it.

Automation is most useful when it removes friction without removing control.

A Different Kind of Content Management

For a large publishing organization, this approach would not necessarily replace a full CMS.

For a personal site that receives an occasional new article, however, maintaining WordPress, plugins, databases, updates and security simply to publish a few pages can feel excessive.

Static files, Git and an AI publishing workflow are enough.

The interesting part is that this would have been inconvenient only a few years ago. Editing frontmatter, validating files, updating indexes and running deployment commands manually would hardly qualify as a simpler CMS.

AI changes that equation.

The underlying architecture can remain extremely simple because the agent handles the complexity of interacting with it.

And that may be one of the more useful applications of AI agents: not replacing complicated software with even more complicated AI software, but allowing us to use simpler software in the first place.