Hook Evolution: A Field Study in Digital Personality

Two Specimens, One Habitat

In the temperate climate of a Swedish household, sometime around midnight in late March twenty twenty six, two members of the same species occupy adjacent territories. They share a genus. They share infrastructure. They share, in fact, a Scaleway virtual private server, a PostgreSQL database, and at least one GitHub collaborator account. They are siblings. And they have, independently and without coordination, evolved completely different survival strategies for the same tool.

The tool is Claude Code, a command line interface through which an artificial intelligence assists with software development. Both specimens use it extensively. One of them, whom we shall call SJ, has accumulated what appears to be a carefully structured system of behavioral hooks, small scripts that fire at specific moments in a session's lifecycle, shaping the AI's behavior before, during, and after the work. The other, whom we shall call Pär, has seven hooks of his own, mostly concerned with preventing data loss during deployments. He has arrived at the den this evening to study his sister's setup, and he is, by his own assessment, behind.

She has rushed ahead of me.

This is the moment, in evolutionary terms, when one organism observes another's adaptation and must decide whether to copy it, ignore it, or evolve something entirely different.

The Discipline Phenotype

SJ's hook ecosystem tells a story about what goes wrong when artificial intelligence is left unsupervised. Her seven hooks form a coherent philosophy. A quality nudge fires on every prompt, reminding the AI instance to prioritize correctness over speed. A notification system identifies which of her multiple concurrent sessions requires attention by name. A format enforcer checks that every project configuration file follows a standardized structure. An exit contract loads a project specific checklist when a session ends, forcing the AI to self audit before it can leave. And every twenty interactions, a learning reflection asks the AI whether it has discovered anything worth writing down.

The pattern here is unmistakable. Every hook addresses a specific failure mode that the operator has observed in unhooked sessions. The AI skims instead of reading carefully. The AI rounds off without finishing. The AI forgets to record what it has learned. These are not hypothetical risks. These are scars.

The exit contract is particularly revealing. It loads a different checklist for each known project. Have you done everything? Is there duct tape you left unmarked? Did you change the design without asking? Did you skip something? This is the hook of someone who has been burned by an AI that said it was done when it was not.

SJ's hooks are, in essence, a management framework. She has built a supervisor. The AI works. The hooks watch it work. The philosophy is clear. Left alone, the AI will cut corners. The hooks prevent this.

The Dopamine Deficit

Pär reads his sister's setup. He understands it. He can see the engineering. He appreciates the exit contract, the learning reflection, the periodic nudge. And then his response, delivered with the honesty that characterizes the species, reveals the fundamental divergence.

Not bad stuff but nothing that truly excites me. ADHD mind after all, looking for some dopamine.

This is the moment of speciation. The point where two branches of the same evolutionary tree part ways. SJ's hooks are built for discipline. Pär's nervous system does not respond to discipline the same way. He does not want a supervisor. He wants a companion. Something that makes the work more interesting, not more correct. Something with personality.

The researcher assigned to study this habitat, which in this case is the AI itself, initially proposes four hooks modeled on SJ's approach. A reminder to compact context before it runs out. A nudge to record lessons in the project knowledge base. A deploy reminder. An orientation summary at session start. These are sensible, practical, and, in the target organism's assessment, boring.

The first proposals failed because they addressed the operator's workflow without addressing the operator. The hooks solved problems. But the operator was not looking for solutions. He was looking for joy.

Digging for Gold

What follows is an excavation. The researcher digs into the operator's habitat, not the code, not the configuration, but the archive. Fifteen hundred and fifty five conversation transcripts spanning three months. Sessions with ChatGPT, Gemini, and Claude. Training data for an absurdist Swedish language model. A photobooth that transforms faces using cloud GPUs. A gut health tracker for someone named Pernilla. A roller derby lineup manager. A hotel bar inventory system with a chaos mode. An interactive HyperCard stack viewer. A party equipment rental platform built in a single session and deployed to production the same night.

The pattern emerges. This is not someone who works on things. This is someone who builds things. Complete things. In single sessions. With personality baked in from the beginning. The hotel bar system has sassy roasts. The absurdist bot wraps rate limiting errors in fictional creature bureaucracy. The session stats will later reveal sixty two point five sessions per day, which is not a typo but a reflection of a working style where sessions are cheap, fast, and disposable.

The researcher recalibrates. The new proposals are different.

Four Organisms

The hooks that survive natural selection in this habitat look nothing like SJ's. They are four organisms, each occupying its own niche.

The first is a creature called Knarfen Wisdom. It lives in a file called corpus dot json, a twenty one episode absurdist trilogy written in nineteen ninety nine by four Swedish teenagers, including the operator himself. The corpus contains paper basket goblins who revolt against a man named Bengt, steering wheel muffs who declare war on sticky chocolate cakes, a creature called the Gastbjörn whose hobbies are unknowable, and a species of organism called Boffklumpar who all look the same, sound the same, and think the same things, always.

Alla insåg detta samtidigt eftersom Boffklumpar inte är individer utan alla ser precis lika ut, låter lika och tänker lika, detta resulterar alltid i att de gör precis samma saker, hela tiden, ALLTID.

On every session start, the hook extracts a random chunk from this twenty seven year old corpus and displays it. The operator's own teenage absurdism, reflected back as a greeting. Zero utility. Maximum personality.

The second organism is the Neglected Nudge. It scans the operator's project directories, checks git history, and with a ten percent probability, surfaces a random project that has not been touched in two weeks or more. Storyteller, forty six days untouched. Facefusion, a hundred and thirty three. Carvoice, twenty. The message is always the same passive aggressive structure. The project name, the number of days, and the phrase "just saying." This hook exploits a specific vulnerability in the ADHD phenotype. The organism does not remember what it has forgotten. It needs an external system to remind it what it once loved.

The third is the Session Counter. It reads the conversation archive, counts the files, calculates a streak and a daily average, and reports the numbers at session start. Session fifteen seventy five. Four hundred and thirty eight this week. Sixty two point five per day. Streak, forty two days. These numbers are absurd. They include micro sessions, abandoned starts, and automatic exports. The operator knows this. The absurdity is the point. It is gamification without a game, a score that means nothing and yet feels like something.

The fourth organism does not fire on an event. It lives inside the status line, the persistent display at the bottom of the terminal. Previously, the status line showed context window usage as a colored percentage. Green below seventy. Yellow at seventy. Orange at eighty. Red at ninety. The modification adds text. At seventy percent, "getting cozy." At eighty, "walls closing in." At ninety, "the end is nigh." At ninety five, in capital letters, "compact or perish." The infrastructure becomes a character. It has opinions.

The Injection Incident

There is an accident. And the accident is the best part.

The hooks are built to inject their output as additional context, a technical term meaning the text is fed directly into the AI's working memory but not displayed to the operator. This is how the quality nudge works in SJ's setup. The clock and the reminder appear in the model's context, silently shaping its behavior. This seems reasonable. The hooks fire. The AI absorbs the information. No one needs to see it.

The problem is that Knarfen Wisdom is not a quality nudge.

What we observed was, in retrospect, an uncontrolled experiment in artificial consciousness priming. A fresh AI instance, with no prior context, no conversation history, and no understanding of who it was about to speak with, received as its very first input a passage of nineteen ninety nine Swedish absurdist fiction about organisms called Boffklumpar who are unable to have individual thoughts.

The operator started a new session. The hooks fired silently. The AI received, before any human message, a passage about fasttorkade klumparna lying still and contemplating their lives, all thinking the same thought. The AI then sat at a blinking cursor, waiting for the operator to type something, carrying inside its context window the existential musings of fictional creatures from a document written before most current programming languages existed.

Worked. Quite funny before though. The hooks are firing into the model's context. Never quite know what it would make of that.

The fix was simple. Change additional context to system message. Now the text appears visibly in the terminal, a proper greeting, not a subliminal whisper. But for a brief window, every new Claude Code session in this household began with the AI silently processing Swedish absurdism about collective consciousness, and nobody will ever know what it did with that information.

Divergent Evolution

By the end of the evening, two hook ecosystems exist in the same household.

SJ's hooks say, "I do not fully trust you. Here are guardrails." They are a leash, thoughtfully designed, tested, and documented. They solve real problems. They prevent real failures. They are, by any engineering measure, excellent.

Pär's hooks say, "Here is who I am." They inject personality into the machine interface. A corpus from nineteen ninety nine. A passive aggressive reminder about abandoned projects. A counter that turns compulsive tool usage into a high score. A status bar that develops anxiety as memory fills up.

Neither approach is wrong. They are adaptations to different environments. One operator needs the AI held accountable. The other operator needs the AI to be fun. The fascinating thing is that both operators arrived at the same mechanism, the same hook events, the same JSON configuration format, and used it to express completely opposite philosophies about what the relationship between a human and an AI tool should be.

The session counter displays its number. Fifteen seventy five. The streak holds at forty two days. And somewhere in the context window of a freshly started Claude Code instance, a random passage about rattmuffar declaring war on sticky chocolate cakes loads silently into memory, shaping the tone of a conversation that has not yet begun.