Your own work is the best training data you’ll ever have. I have a Claude chat that holds 18 months of my own emails. I don’t draft in it. It’s a reference, not a workspace. I query it from other chats where the actual writing and thinking happens.
Advice about AI right now is about prompts. Better templates, better phrasings, better frameworks, all of which are crucial. But it doesn’t address the challenge of AI getting to know you. For example, knowing your career history, background, passions, writing prose, accomplishments, and goals. Because AI doesn’t have that information, it can write a perfectly worded prompt but lack personality and your color.
Let’s focus on training AI on the areas where you communicate most. For example:
- Emails
- Briefings
- Board commentary
- Last RFP responses
- The decisions you’ve made and the words you used to make them
So I built a workflow around it. One Claude chat holds the material as a reference library. I reference that library from new chats, when I’m drafting an RFP, for example. The separation matters because the corpus is too big to live inside a working conversation.
What I figured out across a few passes is that how you ask matters more than which AI you use. The first time I asked the corpus what my voice sounded like, I learned that the patterns I thought defined my writing were less distinctive than I assumed, and the patterns I hadn’t noticed were stronger. By the third pass I was asking better questions, because I’d learned which ones returned real signal and which returned generic best-practice content. The structure I use every time has one section that does most of the work: explicit permission for the AI to disagree with me. Without that, the AI confirms what I already think instead of telling me what’s actually in the corpus. Looking at your own work is supposed to update what you think you know about yourself. That requires telling the AI it’s allowed to push back.
One more thing worth knowing if your corpus is about one topic but you’re writing about something else. Voice transfers across topics; content doesn’t. The patterns that travel are sentence-level: how you phrase a positional claim, the rhythm of your paragraphs, the way you build an argument. The patterns that don’t travel are content-level: vocabulary, technical references, topic-specific framings. When you ask for voice analysis to inform writing in an adjacent domain, you have to be explicit about this, or the AI gives you patterns that sound fluent in the corpus’s topic and fall apart when you try to use them somewhere else.
The corpus matters less than the principle. Pick a body of work that represents how you actually work, drop it into a dedicated chat, and ask it questions using a structure that lets the AI disagree with you. The first pass will be rougher than you expect. The second will start to feel useful. The third is where it compounds.