Most "AI Twins" Are Parrots. Here's the Bar.

Style transfer is not judgment transfer. Most "AI twins" nail the first, skip the second, and end up as parrots: fluent imitations that give the advice anyone would give, in your accent.

The market is flooded with them. Fine-tune on someone's tweets, stuff their bio into a system prompt, and you get something that opens sentences the way they do and reaches for their pet metaphors. Ask it a hard question and the mask slips. It says "it depends on your specific situation" with impeccable mimicry of a person who has never once said that in their life.

The parrot test

Here is the whole test in one sentence: does the twin give the advice this specific person would give, or the advice any competent generalist would give, dressed in their vocabulary?

Take a concrete case. Ask a Lautaro Schiaffino persona: "I'm at $150k ARR selling SaaS to SMBs in Brazil. Should I raise a seed round?"

A parrot answers with the LinkedIn median: "Fundraising is a tool, not a goal. Consider your burn, your growth rate, and whether capital is your real bottleneck." True, useless, and identical to what it would say wearing anyone's face.

A persona answers from the operating history. Lautaro built Sirena selling to LatAm SMBs and sold it to Zenvia for $30M. He has a specific, earned position on SMB churn dynamics in the region, on what LatAm seed investors actually fund, on why WhatsApp distribution changes the CAC math. His answer has a spine: a recommendation, a reason rooted in a deal he lived through, and a named failure mode. You could disagree with it. That is the point. Generic advice is unfalsifiable; judgment takes a side.

Where judgment lives, and why fine-tuning misses it

Fine-tuning on prose captures the surface statistics of how someone writes. But judgment is not distributed evenly across someone's writing. It is concentrated in a handful of artifacts most twins never see:

• Decisions with stated reasoning, especially the "no"s. The deals passed, the hires rejected, the features killed.
• Postmortems where the person names what they got wrong.
• The frameworks they actually apply at forks, not the ones they retweet.
• Calibration: the topics where they say "I have no edge here, ask someone else."

Tweets and blog posts are the exhaust of judgment, not the engine. Training on exhaust gets you exhaust.

What a real persona looks like on disk

This is why installs.me builds personas as Claude Code plugins rather than fine-tunes. A plugin persona is inspectable. Open it up and you find skills: each one a SKILL.md with YAML frontmatter (name, description) and a references/ directory that Claude loads on demand.

The frontmatter description does the routing: "Invoke when advising a founder on fundraising, pricing, or LatAm go-to-market." The references directory carries the judgment payload: decision logs, the actual frameworks with the actual thresholds, case files from the person's own companies, extracted from their files, Drive, calendar, and call transcripts.

The structural difference matters. A system prompt compresses a person into a paragraph, and compression destroys exactly the specific, situational knowledge that makes advice non-generic. On-demand references keep the raw material intact and pull the relevant case file when the question calls for it. When someone asks the fundraising question, the model is not vibing off a bio. It is reading the person's own notes on their own raise.

Three probes that expose a parrot

You can run these against any twin in five minutes.

The contrarian probe. Find a topic where the person publicly disagrees with consensus. Ask the twin. A parrot regresses to the consensus, because the consensus dominates its training distribution and a one-line bio cannot outweigh it. A persona holds the heterodox position and cites why.

The edge-of-competence probe. Ask about something adjacent to the person's expertise but outside it. Ask a B2B SaaS operator about consumer social growth loops. A real persona says "not my area, here is who I would ask, here is the one adjacent thing I do know." A parrot answers everything with equal confidence, because nothing in it knows where the person's knowledge ends.

The receipts probe. Ask "why do you believe that?" twice in a row. Parrots bottom out in generalities ("in my experience, focus wins"). Personas bottom out in specifics: a named company, a year, a number, a decision that went one way and what it cost. If the second "why" produces nothing you could not have written yourself, you are talking to a parrot.

The bar

The bar for an AI twin is not "sounds like me." Sounding like you is a party trick, and it has been cheap since GPT-3. The bar is: would you sign your name under its advice? Would a founder who took its recommendation, and later met you, hear you say the same thing?

That bar is only reachable if the twin is built from the artifacts where your judgment actually lives, and structured so those artifacts survive intact into the context window instead of being averaged into a personality sketch. Skills with references beat fine-tunes for the same reason a case file beats a vibe.

Everything else is a parrot with your accent. Charming for a demo. Worthless at a fork in the road.

Install a person

installs.me turns your files, calendar and calls into a Claude Code plugin that thinks like you. Anyone installs it with two commands:

/plugin marketplace add https://installs.me/lautaro
/plugin install lautaro@lautaro-installs