Inside Anthropic with Dario Amodei #2: The "brilliant friend" spec

Part 2 of 5. Tracing Anthropic's public thinking with FRE|Nxt Labs production commentary.

The source

"Claude can be like a brilliant friend who also has the knowledge of a doctor, lawyer, and financial advisor, who will speak frankly and from a place of genuine care and treat users like intelligent adults capable of deciding what is good for them."

Anthropic's model spec (Claude's Constitution)

Dario Amodei has consistently reinforced this framing in public appearances and in Machines of Loving Grace (entry #3 in this series). The constitution is where it became a design constraint, not just a philosophy.

What we heard

This is a product spec written as a value statement. It defines the failure mode of nearly every AI product on the market: the assistant that hedges, disclaims, and refuses to give a direct answer.

The "brilliant friend" bar is: be as useful to this specific person in this specific moment as a deeply knowledgeable person who actually knows them would be. Notice what that test rules out. The generic answer, the five-bullet framework when someone asked a direct question, the liability disclaimer no one asked for.

If entry #1 (the Responsible Scaling Policy) was the safety floor, this is the product ceiling. The RSP says don't hurt people. The constitution says actually help them.

What we actually do with this

When we build AI advisor products for clients, we run every sampled response through a brilliant friend test:

1. Would a knowledgeable friend say this, or would they say something more specific?
2. Does the response contain any sentence that only exists to protect the AI from liability?
3. Does the response assume the user needs hand-holding they didn't ask for?

If the answer to #2 or #3 is yes, we rewrite the system prompt. Concretely, we audit for:

• "I'm not a licensed professional" boilerplate in responses where it wasn't asked
• "Please consult a human expert" on clear technical questions
• Bullet lists that expand a direct answer into a five-point framework

On the FRE|Nxt Labs AI advisor

Our /advisor endpoint on this site is a live experiment in this. The system prompt is written to answer the specific AI architecture question the user actually asked. Not to explain what RAG is if they already said they're using LangGraph.

We track three signals: % of responses under 150 words (target: 60%+), % of responses containing hedge phrases (target: under 5%), and % of sessions that convert to a discovery call.

The brilliant friend doesn't pad. The brilliant friend answers.

The one thing to steal from this

Run 20 responses from your AI product through this filter: remove every sentence that wouldn't appear in a text message from a knowledgeable friend. What's left should be most of the response. If you lose more than 30% of the content, your system prompt is optimized for legal cover, not for the user.

Next in this series

#3. Machines of Loving Grace (Oct 2024). Dario's 14,000-word utopian essay: the "compressed 21st century," a "country of geniuses in a datacentre," and why the upside case is specific enough to plan against.