AnjadheEvery AI assistant writes in the same voice. Ask any of them for a blog post, a newsletter, a speech, and you get the same tidy, agreeable prose everyone else gets. Competent, and not yours. You spend the next hour rewriting it to sound like you.
Writing Voices fixes that. It is now on for everyone in Anjadhe.
What it is
A voice is a writing style Anjadhe has learned from documents you give it. Usually your own: posts, essays, memos, talks, long emails. It can also be anyone whose writing you have.
When you study a voice, the assistant produces two things, both of which you can see:
- A style guide. One short page describing the voice: tone, sentence rhythm, the words and phrases it reaches for, how it opens and closes, what it never does. You can read it and you can edit it.
- Real passages. Verbatim excerpts from the documents, picked as the most representative. You can pin the ones that carry the voice best and unpin the ones that don't.
After that, any draft the assistant writes in that voice is built from the guide plus those passages plus whatever documents are relevant to the request. The draft names which documents it drew on.
There is no fine-tuning and no hidden profile. What the assistant knows about how you write is a page you can open.
How to use it
- Open Writing Voices from the sidebar, or press Cmd+K and type "voices".
- Create a voice. Name it: My voice, or Mark Twain. It gets
its own folder under
~/Anjadhe/library/. - Add documents. Click Add documents… or drag files onto the page. Posts, essays, books, speeches; .md, .txt, .pdf, .docx and .html all work. A dozen pieces is a good start; more gives a wider sample.
- Press Study. The model reads a spread of the documents and writes the style guide and passages. Read the guide. If a line is wrong, fix it. Your edits stick, even when you add more documents and study again.
- Draft. Press Draft in this voice… and a chat opens already primed for that voice. Ask for the post, the newsletter, the toast. Or in any chat, just say "write this in my voice".
Two things beyond the chat:
- Routines can speak in a voice. A daily review or a morning motivator can be written in a voice you taught. Pick it on the routine's form.
- Try it before teaching your own. Two sample voices are bundled from public-domain works, Mark Twain and Abraham Lincoln. They import with one click, so you can hear what a studied voice sounds like.
If you have a large collection, Study depth in Settings › Reader & Writing Voices sets how much of it one study pass reads.
What you get out of it
Drafts that start where you would have started. The first draft comes back with your rhythm and your moves, not the assistant's. You edit for substance instead of rewriting for sound. For anyone who writes regularly, a newsletter, a weekly update, a blog, this is the difference between an assistant you use and one you gave up on.
A voice you can correct. Because the style guide is a page, the assistant cannot quietly drift into a caricature of you. When it gets something wrong, you change a sentence and it is fixed for good. A fine-tuned model or an opaque profile cannot offer that.
Provenance. Every draft says which of your documents it leaned on. When it borrows a turn of phrase, you can see where it came from.
Other voices, on purpose. Studying a public-domain author, or a writer whose work you own, is a way to study craft. Ask for your own argument in Lincoln's cadence and see what changes.
What stays on your Mac
Your documents are files in a folder you own. To search them by meaning, Anjadhe indexes them with a small embedding model that runs on this Mac, a one-time download of about 330 MB. The index is local, rebuildable, and never syncs. The corpus as a whole never leaves the machine.
If you have pointed the assistant at a model off your Mac, a server you host or your own OpenAI or Anthropic key, then a draft sends what any chat sends: the style guide, the handful of passages chosen for that draft, and your request. Same rule as everything else in Anjadhe. It goes only where you point it, and the LLM Logs show exactly what went.
Draft quality scales with the brain. A local model on a 32 GB Mac does a fair job; a bigger model does a noticeably better one. The guide and the retrieval are identical either way.
A few honest limits
Study reads a sample, not everything. It takes more documents from a bigger corpus and reads from the middle of them as well as the start, but it is a sample.
And a studied voice is an imitation. It gets the rhythm and the moves. It does not have your opinions or your facts. What comes back is a draft that sounds like you, for you to make true.
Writing Voices is in the sidebar. Give it something you wrote and press Study.