AI Content Humanization Techniques

Few-Shot Prompting with Claude Sonnet 4

There’s only one thing that drives me crazy about content creation today.

AI detection tools:)

I’ve tried all the popular humanizers. There are tons of them now because everyone wants this. However, at best, you get readable nonsense.

Maybe these tools work for somebody, but when you’re doing professional B2B content for IT topics and your company (IT outsourcing) lives or dies by SEO, you better be very, very careful before hitting publish on your own or other sites.

I don’t need a bunch of AI detectors. I’m OK with the best one out there: GPTZero.

I love it so much that I pay for premium to check if I sound human. Yes, I constantly blow through the limits, which kicks me over to API billing. But I’m fine with their pricing. No AI humanizer today gets past GPTZero. Only content from media like Entrepreneur gets through.

So why am I even obsessed with humanizing text? Many AI articles are ranking in Google’s top 10 right now

You can get into Google’s top 10 with AI content today. Not just any AI content, but it ranks.

AI-generated articles can sometimes even rank higher in search results than human-written ones.

The problem is that top rankings don’t last for them very long

Give it time and your article drops to page 30-40. More AI articles won’t keep you consistently ranking for different keywords, even short-term.

AI content creation at scale isn’t a sustainable content strategy. You need AI content humanization at scale, too.

Let’s be honest: nobody’s writing without AI anymore. To get visibility on any topic and get clients, you need at least 100 articles, ideally each at 2,000 words.

How long would that take without AI? The math is as follows: one article per day per writer means 100 days before you see results.

Can’t deliver end results (articles that are ready to be published) every single day? What are you even expecting at that pace?

The KPI I shoot for is to write at least one article per writer per day. Use AI if you want, but make it not look AI-generated.

I try to aim for this KPI. But am I actually succeeding at this today

I had my own collection of words for prompts that worked great. Then one day everything stopped.

That sunny August day in 2025, ChatGPT killed all this prompt engineering, and GPTZero decided to eliminate my kind.

A voice from above tried to say something like, “Omnia influit, omnia mutantur. You can’t trick AI anymore. Go work in delivery, bring people food”.

But I was wearing earplugs, so I didn’t hear it.

However, I decided to change how I do things and take it to the next level. That’s when I found Few-Shot Prompting with Claude Sonnet.

How does Few-Shot Prompting work?

Today, I keep a file with a humanization prompt (nothing supernatural, just asking the model to write text like a human).

Most importantly, this file contains before-and-after examples: what AI text looks like and what that same text looks like after processing (this is the version that passes AI detection from GPTZERO).

1. System Prompt […..].
2. Examples
3. Example 1: Input:  […..]. Output:  […..].
4. Example 2: Input:  […..]. Output:  […..].
5. Your Task. Now revise the following text using the same approach shown in the examples above: […..].

This is an endless process:

1. rewrite a text paragraph to 0% or at least 1% on GPTZero,
2. add it to this file.
3. open a new chat,
4. upload this file, and paste the text for transformation

Unfortunately, this is not a silver bullet

Often I have to:

• hit RETRY once or several times
• open a new chat again and repeat the process
• ask the LLM to compare what it did against the examples, find the mistake, and fix it
• take a break and start over (surprisingly, that can help)
• mix text from different attempts, and collecting “human” pieces into one new “green” text,
• make manual edits (funny thing is, if we don’t know how to write, even our manual edits get flagged as AI today – haha!)
• build this few-shot prompting library with examples that satisfy me and feed it to the model all over again (oh god, breaking all the rules of “just a few shots”)
• create new AI-generated text because previous one cannot be humanized (garbage in, garbage out).

In the end, slowly but surely, I’m making progress.

Writing an entire professional text on a new topic, 2,000 words long, in one day, every day, for a long time, is beyond most people’s capabilities on the market.

With AI, it’s possible. And with AI too, sooner or later, I’ll succeed in quickly humanizing text.

This approach has one more downside. The biggest one so far. It only works in Claude (Sonnet is better than Opus). ChatGPT and other AI models are more rigid and do not strictly follow it. If Claude stops working with this method, the whole approach falls apart.

But I am not giving up and I keep building my Few-Shot Prompting library with more before-and-after examples

I want the AI to learn my tone, style, and how I write. I believe the more examples I collect, the better the chance that my AI content humanization tasks will be completed faster and with just one prompt.