The story
I design prompts and workflows that bring AI models to reliable results: editorial copy, images, video, code and analysis. I can also assess rigorously where a model works and where it hallucinates.
In the Linguistics for Digital Communication course, three fellow students and I carried out a study titled “Scalability of semantic intelligence in LLMs”. I was responsible for the software development, the data extraction and the writing of the paper. We put eight models to the test, free and paid versions of ChatGPT, Claude, Gemini and Copilot. The task was to find invented words inserted into 12 real Italian texts.
The result is counterintuitive: paying does not guarantee accuracy. Gemini Pro reached an F1 score of 84.4%, while both versions of Copilot stayed below 50% and also “corrected” things that were not wrong. Today I bring this critical eye to every prompt I write.
How it developed
Daily use of AI in editorial work (ChatGPT, Claude, Gemini, Copilot, DeepSeek) and university research in 2025/2026.
- 2023AI built into editorial workflows for research, drafts and optimisation.
- 2025Prompts for video campaigns, images and competitor analysis (NaturaSnack).
- 2026Study of 8 LLMs, with a scientific paper in CEUR-WS format.
What I can do
- Prompts for text, images and video
- Multi-step AI workflows
- Model evaluation (precision, recall, F1)
- AI usage guidelines for teams
- Generative engine optimisation (GEO)
Tools and keywords
Tools
Keywords
Work and evidence
Co-author of an 8-LLM study: F1 from 42.5% (Copilot) to 84.4% (Gemini Pro).
The study of 8 LLMs
Method, tools developed and results.