If you have spent time building a Custom GPT for email, you probably built a knowledge file with your writing samples and a detailed system prompt. Maybe you uploaded 20-30 of your best emails. Maybe you included style guidelines and worked on the instructions section for a few hours.
And maybe it is pretty good, better than the vanilla ChatGPT output.
But it probably does not fully sound like you. Here is why, and here is what the actual limitation is.
What Custom GPTs actually do
A Custom GPT is a wrapper around GPT-4o with three customization layers:
- A system prompt (instructions you write)
- Optional knowledge files (PDFs, text documents you upload)
- Optional tools (browsing, code interpreter, image generation)
When you upload your writing samples as a knowledge file, the Custom GPT can reference them. But how it uses them is not what most people expect.
The model does not fine-tune on your writing. It does not update its weights based on your emails. It retrieves relevant passages from your uploaded files when they seem useful, and it uses your system prompt instructions to shape the output.
The result is: the model is still GPT-4o, with occasional access to some of your writing as reference material. It has not "learned" your voice in any meaningful sense. It has access to a small sample of it.
The three limitations
1. Sample size is too small
GPT-4o's writing patterns were shaped by training on an enormous amount of text. Your voice is competing with that training by providing a tiny sample: 20 emails vs. the statistical weight of hundreds of billions of tokens.
The model does not transform to match your samples. Your samples are light suggestions against the enormous prior of its training. For generic email styles, this works well enough. For the specific idiosyncratic patterns that make you you (your opener habits, your specific vocabulary, your rhythm), the prior almost always wins.
2. Retrieval is hit-or-miss
Knowledge file retrieval in Custom GPTs uses similarity search. When you write an email, the model looks for passages in your knowledge file that are semantically similar to what you are drafting and uses them as context.
This means the retrieval only works when the email type in front of you matches the email types you uploaded. If you uploaded primarily formal client emails and you need to draft a casual team Slack-length note, the retrieval will not surface the most relevant examples.
3. No cumulative memory
Every conversation starts fresh. The Custom GPT does not remember the email it drafted for you yesterday. It does not know that the prompt from that draft produced output you had to heavily edit. It does not accumulate any understanding of which of its outputs work for you and which do not.
Voice matching requires the model to update its understanding of you over time, based on your actual behavior: which emails you use, which you edit heavily, what the finished sent emails look like. A Custom GPT has no mechanism for any of this.
What voice learning actually requires
Genuine email voice learning requires three things that Custom GPTs do not provide:
1. A large enough sample of your actual sent emails. Not 20 curated samples, but the real distribution of hundreds of emails across different relationship types, contexts, and topics.
2. Embeddings, not retrieval. Your sent emails need to be converted into vector embeddings that can be searched by semantic similarity at draft time. This is fundamentally different from uploading a text file.
3. A feedback loop. The system needs to see what you actually sent vs. what was drafted. When you edit a draft heavily, that edit is information. Over time, a system that processes this feedback learns to generate outputs that require fewer edits.
FinalDraft is built on this architecture. Your sent emails are stored as embeddings. When you draft a new email, the relevant ones surface as context. The system is not a Custom GPT. It is a RAG (retrieval-augmented generation) pipeline trained on your specific correspondence, running inside your email client.
The Custom GPT version of this is fine for occasional use. The embedded version is what closes the gap between "AI-assisted" and "genuinely sounds like you."
Want to see what a voice-trained prompt looks like before committing to a full setup? The Persona Prompt Generator produces a first-person prompt based on your answers. It is the manual version of the insight that the embedding pipeline derives automatically. Free, five minutes, no account.