From Scratch
Train a private alter-ego on your own writing, decks, calls, and emails.
The most useful alter-egos are the ones trained on what you have already done. This guide walks through how to build a private alter-ego from scratch: what data to feed it, how it learns voice and judgment, and what to expect on the first few runs.
A private alter-ego only ever sees the data you give it. It lives in your XO Coworker workspace and is not visible to anyone else.
What feeds an alter-ego
An alter-ego learns from two things: examples of your output, and the surrounding context that shaped it. The richer the second half, the better the first half generalises.
| Source | What it teaches | Good examples |
|---|---|---|
| Long-form writing | Voice, sentence rhythm, what you bother to argue and what you assume | Memos, essays, internal posts, Substack drafts |
| Decks | Structure, framing, the order in which you make a case | Pitch decks, board decks, internal strategy decks |
| Call recordings | How you reason out loud, what you push on, what you let go of | Diligence calls, 1:1s, prospect calls, mentor sessions |
| Email and DMs | How you actually talk, including the parts you would never write in a doc | Sent folder, Slack threads where you were the main author |
| Notes | The half-formed version of how you see things | Daily journals, raw notes app, voice memos |
You do not need all five. You need enough of any one of them for the alter-ego to have something to imitate. As a rule of thumb, fifty pages of your own long-form writing or ten hours of your own talking is plenty to get going.
What it learns, what it does not
An alter-ego learns:
- Your voice. Sentence length, vocabulary, the moves you make when you start a paragraph, the moves you make when you close one.
- Your priorities. What you weight heavily, what you treat as a tax, what you ignore.
- Your shape of argument. Whether you front-load conclusions, whether you build to them, how much hedging you actually do.
- Your taste. The specific kind of clean you are after.
An alter-ego does not learn:
- Facts it was never shown. If it never saw a particular portfolio company in your data, it does not know about that company.
- Live state. If your data was current as of a date, it stays current as of that date until you feed it more.
- Other people's voices, unless those people show up directly in your source material.
The mental model is: it is a version of you with very good notes, not a version of you with a live feed into your head.
Building one
Decide the surface
Pick one piece of work the alter-ego will be responsible for. "You for Memos" works because there is a clear artifact: a memo. "You for Everything" almost always fails because the training signal pulls in too many directions.
Write the surface as a single sentence: "This alter-ego writes my weekly investor update." "This alter-ego drafts replies in the voice I use with portfolio CEOs." "This alter-ego turns my voice memos into board-ready one-pagers."
Gather the source material
Pull together the data that already exists for that surface. For a memo alter-ego, that is your last twenty or thirty memos plus the calls behind a handful of them. For an inbox alter-ego, that is a representative slice of your sent folder.
Drop the files into the alter-ego's workspace folder. Anything text-based (markdown, docs, PDFs, transcripts) is ingested directly. Audio is transcribed first.
Write the brief
Every alter-ego has a short brief that sits on top of the data. The brief is two to five paragraphs in your own words covering:
- Who this alter-ego is and what work it does.
- The voice and stance you want it to hold.
- What it should never do (skip context, hedge into mush, write more than asked).
- The format of its output.
Keep it short. The data does the heavy lifting; the brief points at what to imitate.
Run it on a task you have already done
The first real test is not a new task. It is an old one. Give the alter-ego the same input you had for a memo you wrote three months ago and compare its draft to what you actually shipped. The gap tells you exactly where to add material or tighten the brief.
Iterate on the gap
If the voice is right but the structure is off, the brief needs to be tighter. If the structure is right but the voice is off, you need more long-form examples in the source data. If both are off, the surface is probably too broad: split the alter-ego in two.
A worked example: "You for Memos"
Someone who writes one investment memo a week, has done so for two years, and has roughly a hundred shipped memos.
- Surface: drafts the first version of a memo from a deal note and a call transcript.
- Source: the hundred memos, plus the calls behind the last twenty of them, plus their internal Slack thread where the deal was first discussed.
- Brief: three paragraphs on stance ("skeptical, fast to a verdict, never buries the ask"), format ("six sections, conclusion first, max two pages"), and don'ts ("no hedge phrases, no executive summary, no bullet salad").
- First test: hand it the call transcript and deal note from a memo written six months ago. Compare drafts.
A good outcome is that the first draft would have saved an hour of writing if it had existed at the time. A great outcome is that the alter-ego catches something you missed in the original.
Common pitfalls
- The surface is too broad. "Write all my comms" is not a surface. "Draft replies to portfolio CEOs" is.
- The data is too clean. If you only feed the alter-ego your polished output, it learns to imitate polish, not thinking. Include drafts, notes, and the messy version.
- The brief contradicts the data. If your brief says "be concise" but every memo in the data is three pages, the data wins. Pick one and align them.
- You judge it on the first run. The first run is a calibration. The second and third are where you decide if it is working.
When the alter-ego starts producing drafts you find yourself lightly editing instead of rewriting, you are done with v1. From there it is a question of feeding it new material as the work evolves.