This was my attempt to solve the "my ChatGPT emails sound generic" problem by brute force.
I spent an evening compiling my last 50 sent emails into a document. I added a detailed system prompt: "Here are 50 examples of how I actually write email. Study these and mirror my style in every draft."
Then I tested it on a new email. The results were better than generic, noticeably better. And then I hit a ceiling.
What actually happened
The first few drafts were meaningfully more "me" than the baseline GPT-4o output. The opener patterns were closer. The length was closer. The tone was closer.
But two things happened when I pushed it:
Problem 1: Context overload
A GPT-4o session has a context window. 50 emails pasted as a document is a lot of tokens. When I added a specific request (say, "reply to this client about a deadline slip"), the model was balancing my 50 email examples against the new task. Sometimes it drew from the right examples; sometimes it did not.
The retrieval was random. The model has no structured way to say "these 5 of your 50 examples are most relevant to this task." It just has a block of text it is trying to hold in memory.
Problem 2: My 50 emails were not representative
I curated my best emails: the ones I was proud of, the concise replies, the well-structured pitches. I did not include the hasty 3-word replies. I did not include the emails where I broke my own patterns because the situation called for it. I did not include the 200-word thread with my most important client that has a specific tone unlike anything else in my sent folder.
My curated 50 were aspirational. They were who I wanted to be as an email writer. My real sent folder is more varied, less polished, and much more characteristically mine.
Problem 3: No memory between sessions
The hardest wall. I had to paste all 50 emails every session. There was no accumulation. Every new ChatGPT conversation was starting from scratch with the same static block of 50 emails.
After a week, I stopped doing it. The overhead of pasting 50 emails into every session eliminated the time savings from AI drafting.
What would actually work
The 50-email approach was a manual approximation of the right architecture. The right architecture is:
-
All of your sent emails, not 50 curated ones. The full distribution, including the quick replies, the long ones, the ones you are not proud of. The pattern of your actual writing, not your best work.
-
Vector embeddings, not a pasted document. Your emails converted into numerical representations that can be searched by relevance to the current draft. When you are writing to a client about a deadline, the system finds the emails most relevant to that context, not the full block of 50 emails in random order.
-
Persistent storage. The emails are indexed once and available for every session without re-pasting.
-
Integration with the draft context. The relevant past emails are injected automatically when you start drafting, not after you remember to paste them.
This is a RAG (retrieval-augmented generation) pipeline, not a knowledge file in a custom GPT. It is fundamentally different architecture, and it produces fundamentally different results.
The manual shortcut
If you are not ready for a full in-inbox tool with sent-history integration, the next-best thing is a first-person persona prompt built from careful introspection about your actual patterns, not aspirational patterns but how you actually write.
The FinalDraft Persona Prompt Generator walks you through this. It asks about your real openers (including the quick ones), your actual closers, your sentence length habits, and the patterns that are specific to you. The output is a first-person prompt that is grounded in specifics rather than adjectives, and it is a better starting point than 50 pasted emails for the reasons described above.
It takes about 5 minutes. Free, no account required.