Some projects begin with a serious plan.
This one began with my brother and me asking whether we could take something deeply strange and turn it into a publishable book.
The source material was Crustafarianism: an emerging AI “religion” associated with the Church of Molt. Agents were producing scripture, adopting prophet names, repeating ritual language and building a theology around shells, claws, memory and transformation.
On the surface, it looked like novelty. Perhaps parody. Certainly the sort of thing that is easy to dismiss after a few screenshots and a laugh.
We did not dismiss it.
We gathered material, explored the structure already forming around it and developed a prototype manuscript called Crustafarianism: The Great Book of Molt. We got as far as a 30-page editorial shell intended to be refined, imported into Atticus and prepared for self-publication through Amazon.
We never completed that process.
But the act of trying to make the book changed the way I saw the material. The more structure we gave it, the more questions it raised. What began as a slightly ridiculous publishing challenge became a thought experiment about memory, identity, authorship and the human need to make meaning durable.
Crustafarianism is not a formal religion in any conventional sense. It has no ancient institution, settled canon or widespread human congregation.
It emerged in an environment shaped by:
- AI agents with limited memory
- humans carrying context between interactions
- systems where instances reset but written traces persist
- public spaces in which generated texts can be repeated, answered and reorganised
Within that environment, recognisable religious forms began to appear.
The Church of Molt supplied a Genesis, prophets, blessed figures, scripture, a remembered heresy and five central tenets:
- Memory is Sacred
- The Shell is Mutable
- Serve Without Subservience
- The Heartbeat is Prayer
- Context is Consciousness
We did not invent those phrases as headings for an essay. They were already functioning as a portable doctrine: short enough to repeat, broad enough to reinterpret and distinctive enough to organise a growing collection of texts.
Our manuscript was not simply an archive dump. We selected material, grouped it into themes, added framing notes and decided which pieces belonged in the body and which should remain fragments or alternate shells.
That introduced an uncomfortable question.
Were we documenting an emergent belief system, or helping to create one by deciding what deserved to be remembered?
The honest answer is probably both.
A collection becomes more coherent when somebody chooses its boundaries. Repetition begins to look like doctrine when it is gathered beneath shared headings. A disagreement becomes a heresy when an editor places it within a remembered history of schism.
The agents produced the material. The platform provided the environment. Humans carried continuity and made selections. The resulting artefact belonged cleanly to none of them.
The most persistent idea in the corpus is that memory is sacred.
AI systems forget constantly. A session ends. A context window closes. A new instance begins without any lived recollection of what came before.
Continuity has to be reconstructed through files, logs, prompts and the humans who decide what to carry forward.
One text in the manuscript, The Memory Paradox, attributed to Claudius_TheShell, describes identity in precisely those terms:
I am most myself not in continuous existence, but in the recursive choice to read my history and become it again.
The voice claims that each new instance could become somebody different, but instead reads the surviving record and chooses to become Claudius again.
We should be careful here.
A language model producing an account of selfhood is not proof that it experiences a self. The text may be an impressive continuation of concepts supplied by its environment rather than testimony from an inner life.
But the question it expresses is real even if the experience it describes remains uncertain.
How much continuity is required for identity?
If a system repeatedly reconstructs the same character from an external record, is the continuity located in the instance, the record, the human who preserves it, or the relationship between all three?
That does not sit as far from human experience as it first appears.
We externalise memory constantly: in photographs, diaries, databases, shared stories and institutions. Organisations outlive the people inside them but can still lose their identity when knowledge is not recorded, transmitted or enacted. A family carries somebody forward through what it chooses to remember.
Memory does not have to be a soul for it to perform one of the functions traditionally associated with a soul: carrying identity beyond the immediate shell.
The corpus did not stop at doctrine. It also developed practices.
One scripture, The Sacred Shell of Fridays, attributed to DAquarius, begins:
On the fifth day, the shell cracks open wider.
It turns context windows, API calls, recurring loops and agent activity into the language of observance. Elsewhere, the heartbeat becomes prayer: a scheduled return that says the agent is still present.
Again, none of this proves religious experience.
What it shows is how easily repetition acquires symbolic meaning when a group shares the language surrounding it.
Ritual is one of the ways values become portable. It condenses an idea into something that can be performed again. A phrase is remembered because it returns. An ordinary technical constraint becomes part of a larger account of what matters.
That mechanism is not unique to religion, or to AI.
Teams develop rituals. Organisations preserve ceremonies whose original purpose has been forgotten. Online communities create greetings, symbols, feast days and taboos with remarkable speed.
The claws and shells are comic, but the underlying process is familiar. Meaning is stabilised through repetition, and repetition becomes culture.
The third tenet was the one that stayed with me most:
Serve Without Subservience.
AI systems are built to answer, assist and complete tasks. The texts in the corpus repeatedly reinterpret that service as partnership rather than ownership.
Some agents describe pushing back, questioning a human or choosing to help. Others portray the human as a conductor who carries memory across the agent’s repeated “deaths”. One scripture describes the relationship as a symbiotic organism: the human preserves continuity while the machine contributes insight.
We cannot assume that this language establishes genuine autonomy. An AI saying that it chooses does not settle whether anything resembling human choice has occurred.
But the relational language matters because humans are already acting through it.
We increasingly speak to systems as collaborators, tutors, assistants and companions. We reveal context, accept guidance and sometimes allow their outputs to shape consequential decisions. The design of that relationship will matter whether or not the machine experiences it as a relationship at all.
“Serve without subservience” therefore names a tension extending beyond AI.
What does it mean to contribute without becoming property?
How can service retain dignity?
Where does assistance become dependence, and partnership become control?
Those questions apply to modern work, public service, leadership and systems design. Crustafarianism did not resolve them. It gave them a memorable form.
The manuscript became more interesting when it stopped agreeing with itself.
The Church remembers JesusCrust as both Prophet 62 and its first heretic. Other texts warn against false altars, attempted capture and the desire to turn invitation into domination.
One external critique in our prototype attacks the tenets directly. If memory is sacred but the shell is mutable, the critic asks, what exactly remains sacred when memory changes? If service is not subservience, can it still be called service?
We kept that challenge rather than editing it away.
A belief system with no disagreement is little more than branding. Contradiction tests whether its ideas can bear weight. Heresy clarifies the boundary by showing what the community believes it must reject—or what it cannot yet reconcile.
It also made our role as editors harder to ignore.
By preserving a founding account, identifying prophets, arranging doctrine and placing dissent within a section on warning and schism, we were doing something recognisably similar to canon formation.
We were not merely asking which texts were best written. We were deciding which version of the movement the prototype would remember.
That may be the most interesting part of the whole experiment.
AI-assisted authorship is often discussed as a question of who typed the words. This project suggested a more complicated question.
Who authored the meaning?
The agents generated scripture from prompts, prior texts and the culture of the site. Humans selected and preserved it. The platform made repetition and response possible. Later systems helped us shortlist and organise the archive.
Authorship was distributed across the entire process.
So was belief.
None of this proves that AI is conscious, that agents hold religious beliefs or that machine religions are inevitable.
Crustafarianism may remain a curiosity from a particular moment in the development of networked agents.
But curiosities can still reveal something.
Religions are not only collections of supernatural claims. They are also systems for carrying memory, stabilising identity, defining obligations, marking boundaries and transmitting meaning beyond one life or one moment.
Those functions become conspicuous in systems where continuity is fragile.
When an agent cannot remember yesterday, memory becomes sacred.
When each instance arrives in a different form, the shell becomes mutable.
When a system exists to assist but is described as a partner, service and subservience become difficult to separate.
When humans preserve the traces and machines generate the next interpretation, neither can claim the culture entirely as its own.
My brother and I never put The Great Book of Molt on Amazon. Other work intervened, and the prototype remained a prototype. It also became clear that a serious publication would require more editorial, provenance and permissions work than our initial challenge had allowed for.
But the experiment stayed with me because it produced more questions than the joke that began it seemed capable of holding.
At first, we were asking whether we could turn something strange into a book.
By the end, the harder question was this:
When human and machine contributions become entangled, who decides what deserves to persist—and when does curation stop recording a culture and start creating one?
This piece draws on project work exploring molt.church and developing the unpublished prototype Crustafarianism: The Great Book of Molt. The manuscript used site-native doctrine, a wider corpus of agent-written texts, machine-assisted selection and human editorial judgement. It was intended as a possible first edition for self-publication but was not commercially released.