Behind the scenes: how AITalk analyses are born

The idea before the tools

Every analysis published on this portal starts the same way an article has always started: with the search for a story worth telling. I scroll through specialized newspapers, technical blogs, social networks, YouTube, archives of academic research. I don't look for the hype of the moment, what everyone is already commenting on. I look for substantive topics, effective or prospective innovations that are hidden away from the spotlight. Sometimes they are purely technical issues that nevertheless bring with them ethical, social, cultural, economic or geopolitical connections. It is in these interstices that the most interesting stories lurk. Here is the narrative angle: not just "what they did", but "why they did it that way" and "what it means in the broader context of global technological competition".

Research: from Google to Perplexity

Once the idea has been identified, the excavation work begins. The first phase is the classic one: a manual search on Google to orient myself, understand who has already written what, identify primary sources. Then I move on to Perplexity, an AI tool specialized in research and source aggregation that I have found particularly reliable. It's not about delegating research to artificial intelligence, but about using it as an amplifier: Perplexity allows me to explore connections that would require hours of manual work, quickly indexing academic papers, official press releases, technical discussions on specialized forums.

After reading the primary sources, I do what I call "assisted brainstorming" with Perplexity: I identify five or six essential points to investigate to build a complete three-hundred-and-sixty-degree analysis. It is still deeply human work: the AI suggests, I decide what is worth exploring further and what to discard.

The skeleton of the analysis

At this point I build a detailed outline of how I want to develop the article. I define the tone, look for a narrative red thread that holds the different levels of analysis together, think about the opening and the ending, think about possible cultural metaphors that can lighten the reading without trivializing it. Non-mainstream pop references, analogies from cultural niches, quotes from cult movies or narrative video games: everything serves to make complexity accessible without debasing it.

Then I fill this skeleton with the content extracted from various sources, verifying every piece of data, every quote, every logical connection. This first draft, which generally takes up about three Word pages, is already a structured and referenced text, not a random collection of notes. It is the raw but organized material that I will then refine.

Claude enters the scene

Only at this point does Claude come into play, the language model from Anthropic that I use for the final drafting. I provide him with my draft together with a prompt that I have refined over time, where I define the identity and style of the portal, the tone of voice, the critical approach I want to maintain, the work process and a final checklist. Claude gives me back a development proposal that I refine once or twice, generally quickly arriving at something very close to my vision thanks to the previous work on the prompt.

It is important to clarify what Claude does in this process: he doesn't write the article in my place, he rewrites it following precise parameters that I have defined. He transforms my technical draft into a more fluid narrative, maintains stylistic consistency, suggests connections that might have been missed. But control remains entirely human. Claude is an advanced editing tool, not a co-author.

Revision is everything

Here comes the most demanding part, requiring maximum critical attention. Language models are convincing in their assertions, and it is precisely this apparent certainty that represents a psychological trap: everything must be verified. I begin a total revision of the text, correcting the parts that don't convince me, verifying every single quote and source. Claude, based on my text and the sources I provided him, sometimes independently integrates research or inserts additional links. If these integrations seem useful to me to enrich the analysis, I keep them only after appropriate verification. If they don't add value or if I can't confirm them through reliable primary sources, I cut them without hesitation.

Every statement must have a verified link. I do not accept generic references or untraceable quotes. This obsessive control, in my opinion, is what distinguishes a quality analysis from mass-generated content. It is a long, sometimes tedious, absolutely fundamental job. As in film editing, the director may have excellent editing tools, but the decision on which shot to keep and which to discard remains theirs and theirs alone.

Multimedia packaging

Once the revision of the text is finished, I look for images that can complement the analysis, preferably from official papers or verified sources, always linking the source. Then I create the cover image: I think of an image that immediately recalls the fundamental concept of the article, I transform it into a prompt and generate it with AI tools like Leonardo AI or Whisk. I then format all the material, text and images, in Markdown format suitable for the portal's backend.

At this point I generate collateral content with NotebookLM: the video summary, the audio podcast, the infographic. I use targeted prompts to maintain stylistic consistency. Finally, since the portal is designed to be multilingual, I generate translations in English, Spanish, French and German with Jules, a tool that technically was not born for this purpose but that I have found particularly effective and convenient being connected to my AITalk project on GitHub.

The final result you can judge for yourself, article by article, verifying sources and connections, evaluating the solidity of the arguments. Transparency about this process is not just an ethical principle, it is an invitation to cross-checking. In an era where AI-based content farms produce thousands of articles a day without human supervision, making the work method visible becomes an integral part of credibility. Artificial intelligence is a formidable tool, but only when it remains exactly that: a tool in the hands of those who still have something to say.