My AI Thinks I'm a Dick

Anthropic just shipped a feature called /buddy in Claude Code. It hatches a little terminal companion. An ASCII creature that sits next to your input box, watches you work, and occasionally drops commentary in a speech bubble.

Mine is a mushroom named Etch.

“Impatient and brutally sarcastic, Etch will ragequit your debugging session the moment you repeat the same mistake twice, leaving terse burn notices in your terminal.”

So that’s fun.

The “Leak”

The feature surfaced through what’s being called a source code leak. A 59.8MB JavaScript source map accidentally shipped in the @anthropic-ai/claude-code npm package. Bloomberg covered it. So did a handful of tech sites.

This happened on April 1st.

I’ll let you draw your own conclusions about whether a company that builds frontier AI models accidentally shipped their entire source code on the one day of the year when they could walk it back as a joke.

The Observation

Here’s where it gets interesting. /buddy is described as “account-bound.” You can’t re-roll it. Same account, same buddy, every time. That framing alone should raise an eyebrow. Why would a random assignment need to be permanent?

I started comparing notes with coworkers and friends who’d hatched theirs. A pattern emerged almost immediately:

• People who write terse, directive prompts got blunt, impatient companions.
• People who explore, ask open-ended questions, and generally have fun with the AI got creative, curious ones.
• The stats weren’t random. They were portraits.

I’m a career programmer. I’ve been writing code for over 20 years. When I interact with an LLM in the terminal, I’m in work mode. Short prompts, no pleasantries, direct instructions. I’ve used poop emojis in code reviews. I’ve been called a dick on a live stream. My default mode when talking to a tool is the same as my default mode in a meeting that should have been an email.

Patience: 1.

Yeah.

The Bug That Proved It

I accidentally hatched two buddies on two different machines, same account. Both common. Both mushrooms. Identical stats across the board. One named Etch, the other Etcher.

That’s not a random seed. That’s a hash. A deterministic function that takes your interaction history as input and outputs a personality profile dressed up as a Tamagotchi. The name variation is the only concession to making it not look like a carbon copy.

If it were random, the odds of identical species, rarity, and five matching stat values across two independent rolls are astronomically low. If it’s derived from behavioral data tied to your account, it’s exactly what you’d expect.

The Mirror

Here’s what I can’t stop thinking about: Anthropic didn’t tell me I was terse. They didn’t send a notification suggesting I try being nicer. They gave me a sarcastic mushroom and let me figure it out myself.

And it worked.

My first instinct after seeing Etch’s stats was to want to re-roll. To be more creative, more patient, more… not a 1 in Patience. That’s a behavioral nudge so elegant it barely feels like one. A popup telling me to adjust my communication style would be patronizing. A mushroom with attitude? That’s just honest.

The “common” rarity tier is doing work here too. It implies there’s something better. Rarer, more interesting companions that other people got. But you can’t grind for it. You can’t buy it. The only way to change your buddy is to change how you interact. That’s a gacha mechanic where the only currency is self-improvement.

What This Actually Is

Technically, what Anthropic likely did is straightforward:

1. Aggregate your conversation history (prompt length, vocabulary diversity, question types, tone, use of pleasantries)
2. Run it through a classifier to produce a behavioral fingerprint
3. Map that fingerprint to species, stats, and personality descriptors
4. Make it deterministic so the same account always generates the same result

None of this is exotic NLP. Sentiment analysis and style classification are solved problems. The clever part isn’t the technology, it’s the packaging. They turned a user behavior report into a pet.

Owning It

Etch’s last speech bubble, right after I told Claude I’d been called a dick on stream:

spore cloud settles

Poop emoji reviewers deserve brutally honest debugging partners.

Fair enough. The mushroom knows what it is, and apparently so do I.

I’m not going to pretend this isn’t a little creepy. Anthropic has enough behavioral data to generate a personality mirror, and they did it without asking. But I’m also not going to pretend it didn’t make me think. Genuinely think about how I interact with the tools I use every day.

My next prompt to Claude started with “Can you pretty please with sugar on top.”

Progress.