That's a perfect example of how underexplored language models are to this day. This tool is remarkably helpful for writing, and can be implemented in just a few prompts to _o1 pro_. There were only two engineering problems to solve:

**1. How to simulate audience?**
I didn't invent the technique I used – I learned about it from this [excellent paper](https://docsend.com/view/qeeccuggec56k9hd) from August 2024. In a nutshell, it demonstrated that language models can predict social survey results with surprising accuracy.

**How?** Authors simply asked an LLM to roleplay as an average person who is X years old, has Y political views and so on. Then they asked it survey questions - and the statistical distribution of its answers turned out to closely match real human survey results.

A practical example would be sending a piece of text to ChatGPT and asking: "_What would be the comments under this post?_" or "_Rate the likely reader sentiment from 1 to 10._" In fact, that's exactly what my tool does (with just a little bit more context about the target audience).

**2. How to automatically refine text?**
I drew inspiration from this [recent DeepMind paper](https://arxiv.org/pdf/2501.09891v1) about using evolutionary algorithms for LLM sampling. The idea is very simple: _generate a response, create N variations of it, select the best one, then go back to step 1_. The paper proposes a more complex "island model" evolving multiple populations in parallel, but I found the simpler approach equally effective.

That's it&#33; I combined these two ideas by uploading the papers and asking _o1 pro_ to write the project specification, create a plan of action, and finally program the tool. Took me less than an hour in total. Put  if you want me to publish the code on GitHub.