Listen
Audio placeholder · 10 min
No file in v0. The bar at the bottom is a stub. Feeds live on Listen when they exist.
The Brief
Forward this. Then decide if the essay is worth your time.
- 01
0:00 · Twin vs doll
A twin retrieves and cites. A doll improvises in your voice.
- 02
0:03 · Guards are the product
Scope, secrets, and 'do not send' are not extras. They are why the twin is safe to run.
- 03
0:06 · Provenance or it did not happen
If a claim cannot point at a note, a file, or a date, treat it as fiction.
- 04
0:09 · Public vs private
The public brand is not the private vault. Do not wire them together for a demo.
Essay
A useful digital twin is a grounded proxy of your judgment: cited to real work, scoped by hard limits, and replaceable at the harness layer. A digital doll is a persona with your name on it. Dolls impress in demos. Twins survive Tuesday.
I sat in a room where an executive talked about digital twins for leaders, not demo reels. The line that stuck was simple: every task, every role, every day. I had already been building a personal twin for more than two years. That conversation was a catalyst, not the start.
The mistake I see builders make is dressing a chat window in autobiography and calling it a twin. That is cosplay.
Twin vs doll
Doll: "You are Ruby. Speak as Ruby. You know everything about Ruby."
Twin: "Answer about this person with a citation to a dated note. If you cannot cite, say so. Do not predict what they would do. Do not speak as them for outbound work."
The difference is epistemology. Dolls invent continuity. Twins admit gaps.
Research in the field keeps showing the same uncomfortable result: dumping rich personal detail into a simulator barely moves accuracy over knowing age and job. A simulator has no way to know when it is wrong. A citation does.
So I treat the twin as:
- Judgment and voice captured from real work already done
- A model that is only the mirror
- Context and controls that make the mirror useful
If Copilot only knew you from this week's emails, would you be proud of what it learned? That answer tells you where to start.
Build from your best day, not the noise
Speed without a standard just produces average faster. Tools amplify whatever you feed them.
Before I open an agent on a recurring task, I write five lines:
- The outcome I need
- Who feels it
- What good sounds like
- What must never happen
- What done means
Then I run the same task tomorrow and edit the five lines when the draft misses. That is the whole method. Repetition beats inspiration. The twin improves because you fix the brief, not because you chased a new model on Friday.
Public examples I actually use:
- RivaFlow: session notes and training loops as structured judgment, not vibes about "being a coach"
- Photography (The Room): house style and proof pipelines, so agents cut and caption without inventing customers
- Personal harness: vault as memory of record; heads (Sage, Super Doer, Cursor) plug in and can be swapped
Guardrails that make it a twin
A doll has personality. A twin has a charter.
Mine, in public terms:
- Vault is durable. Head is replaceable. If a chat product dies, the notes and contracts remain.
- Grounded proxy, not ventriloquism. Cite dated artefacts. Do not answer as me for stakeholder mail.
- Scope out of modelling. Some life domains stay lived, not extracted into digests and claim files.
- You own the send. Draft in voice, flag generic AI, ask one hard question, edit, then human clicks send.
- Prose is not a control. Rules in markdown are wishes until code and evals refuse the bad copy.
That last one matters. I watched tools that never read the policy file cheerfully promote private journal text into "evergreen candidates." They passed every check they had. The fix was a scope-guard in code plus evals that fail the build. Twin work without enforcement is cosplay with better lighting.
SMB to enterprise ladder
Solo / SMB: Pick one recurring task. Save the five lines. Keep a dated folder of "best day" samples. Require the agent to quote which sample it used. Draft-only for anything external.
Scale-up: Role charters as runbooks (outcome, audience, must-not, done). Shared examples library. Citation required in PR descriptions and customer replies drafted by agents.
Enterprise: Twin doctrine in the operating model: what may be modelled, what must be cited, what never leaves the journal path, who owns send, how evals gate promotion from draft to durable claim. Treat "AI persona of the exec" demos as dolls until those controls exist.
Failure modes
- Speaking as the human in outbound channels
- Training on this week's noise instead of best-day artefacts
- Expanding scope because the model "feels intimate"
- Swapping the vault for whichever chat remembers the thread
- Confusing a Friday wow with a system that works at peak on the second Tuesday
Closing
A digital doll is easy to ship and hard to trust. A digital twin is slower to stand up and boring in the best way: citations, charters, evals, replaceable heads.
Build the twin from your best day. Guard what must stay lived. Own the send. Leave the cosplay for Halloween.
Related
Want the next edition? Join the newsletter placeholder. No modal. No auto-send.
Ruby Wolff
Gem Alpha. Shipping in public. Builder notes, not a news desk.