A Fingerprint of How You Write, Make AI Sound Like You

I built a tool that reads a person’s own documents and analyzes how they write. It turns that into a stylistic fingerprint that any AI model can follow, so the output sounds like them. This is important when you have an automated document creation process, where AI writes new content, adds to what you already have, or reviews it.

The problem

Many clients want to automate their document writing process, whether that is a proposal or a press release. Even when they write everything in their own words, or use AI to augment it, or rewrite that content, it starts to sound like AI. That also happens when they take an existing document and have AI create a new one from it, a derivative. This often means they have to heavily edit it, which costs more time.

Every single time an LLM touches the work, even original work, it changes the wording and makes it sound like AI. No matter how many times you tell an LLM to write like you, even with examples, even with the recommendations from the web on how to make AI sound more like a person, it still reads like an LLM wrote it.

What I built

To solve this, I built a tool that generates a style fingerprint. That fingerprint gives an LLM the key characteristics and attributes of your style, so it outputs content that is more like you.

Whether you are creating derivative content or something new in an automated way, this fingerprint shapes content to sound more like you than like AI. It will not be 100% perfect. Nothing is.

This fingerprint can map an individual’s style, or your firm or brand style.

The tool creates the style fingerprint by analyzing several documents you have really written, then tunes it by preference learning, where you select from a list of fragments and pick which one sounds more like you. That analysis uses academic techniques, and my own techniques, to create a stylistic fingerprint.

How it works

The stylistic fingerprint captures the kinds of things that make up how you write. Some of these you never thought about while you were writing. It captures how long your sentences run, and where you drop a short one in. Whether you write “it is” or “it’s”. Which verbs you use frequently. How you use the small words you never noticed, “the”, “to”, “and”, “but”, or “because”. This fingerprint holds 22 attributes that are specific to you and how you write.

The tool uses math to analyze the writing and create the characteristics and attributes across multiple dimensions. You provide samples of real documents you have written, and it analyzes them and finds the patterns across them.

For example, the tool uses Burrows’s Delta, a technique from forensic linguistics. It also measures the rhythm, shape and construction of how you write, and creates archetypes of the language you use.

In addition, you are able to fine tune it with options of phrases, where you pick the ones that sound most like you. This helps refine the fingerprint.

The style fingerprint is content type agnostic, and works for whatever content type you are asking an LLM to generate.

To use it you provide the fingerprint to an LLM, and instruct the LLM to use it, along with the content you want to use. This can be applied as part of a workflow, or after the content is generated.

See it

The same document, given to the same model twice. Once on its own, once with the fingerprint.

Without the fingerprint“In today’s fast-paced business environment, automation has become a game-changer for small businesses looking to streamline their operations. However, it is important to note that successful automation begins with a deep understanding of your team’s unique processes. By taking the time to capture how your team thinks, you can ensure that any solution you implement is both effective and sustainable. Ultimately, this approach empowers your business to unlock its full potential.”

With the fingerprint“Most of what matters in a process is never written down. It is just what the team does, and it costs them time every week because no one else can do it. So before I automate anything, I interview the people who do the work and capture how they decide, where they make tradeoffs, and what a good result looks like to them. The automation gets built from that, so it still works on the day the person who normally does it is out.”

I built this for my practice and for my clients, so the output of their content workflows sounds like them, and so they can use it inside their organization or inside the workflow itself. I am not selling the tool.

Back to Portfolio