Minerva is a tool to empower the creative process.
We draw, gesture, hum, sketch, annotate, and associate. We speak in metaphor and shorthand. We make connections that are oblique but exact. Minerva is built on the conviction that a tool for creative work should recognize all of this, and that artificial intelligence becomes genuinely useful only when it is placed in service of the user's own way of thinking, not substituted for it.
Philosophy
Minerva should help exercise critical thinking, promote wisdom, offer overall perspective, and understand and validate many personal forms of communication.
We are living through a moment when artificial intelligence is taking over many tasks and products, fueled by rapid development and the constant promise of something better just around the corner. That pace has produced a proliferation of products and services that often feel like recycled ideas, shaped by how the systems behind them are built, tools that are endemically designed to influence users to prioritize computing efficiency rather than listening to them. What gets lost in that rush is something fundamental: that human beings do not think, communicate, or create in a single register. If the user is to be at the center of the generative process, artificial intelligence should be a seamless companion along the journey.
For starters, the way we interact with digital tools has been text-heavy. Text is simple for computers to understand, but it overlooks the many other ways humans communicate: diagrams, voice, pictures, gestures, moods, second meanings, poetry, drawings and sketches, shorthand, music and melodies, onomatopoeia, and references that may not be directly connected to their source yet still evoke an idea or emotion that suddenly clicks. Even the way we speak varies across languages, significantly changing meaning and communication style.
Several intellectual traditions inform this philosophy. Noam Chomsky's work in linguistics suggests that language is not merely a learned behavior but a deep cognitive structure, that the capacity for meaning runs beneath the surface of any particular tongue or medium. John Berger's writing on perception, particularly in Ways of Seeing, argues that we understand the world through images and visual experience long before we reach for words, and that much of what we know resists verbal translation entirely. This way of thinking points toward the same essential truth: human communication is layered, embodied, and irreducibly plural. A tool that only reads text is not reading people, it is reading a narrow transcript of them.
The creative process is as complex as understanding consciousness, an endless quest, but one common observation holds: it involves continually searching the memories and experiences saved and categorized in the vast archive of the brain. We may find them directly or, more importantly, unconsciously mix and cross-reference them so ideas spark in response to an assignment. From Minerva's point of view, artificial intelligence is a tool that can help retrieve and cross-reference these materials more effectively, and in turn empower an individual or collective creative process.
In practice
This tool's philosophy takes a particular shape in the product. These four levels of interactivity are designed to structure the user's creative process without input constraints:
The Projects Collection
The Projects Collection is the first thing a user sees: a personal archive of every project they have created or been invited into. Each project appears as a named card showing its last-modified state, its type, and a thumbnail of its Field. From here, a user can open an existing project, create a new one, or organize their work into groups. The Collection is not a dashboard of activity or a feed of notifications, it is a quiet, spatial index of everything the user is thinking about. It exists to make the full scope of a person's work visible at a glance, and to make moving between projects frictionless.
The Dashboard
A project in Minerva is a constellation of cards on a spatial canvas, the Field. Each card holds a piece of the work: a chapter of prose, a reference that supplies context, or a note that captures a passing idea. The relationships between cards are drawn explicitly. A solid line means one piece of work contains or leads to another; a dotted line means one piece of work informs another's reasoning. The shape of the project becomes legible at a glance. Moving a card moves a part of the argument. Drawing a connection makes a relationship the AI will know about and reason from.
The Editor
A chapter opens into an editor whose margins are as active as its body. Highlights carry semantic weight, yellow for what matters, red for what should be cut, green for what should grow. Margin notes carry questions, second-guesses, reminders to come back. Comments carry conversations the writer holds with themselves. When the AI is asked to analyze the chapter, it reads all of these as signal: the writer's own marks on the work become the instructions for what to do with it.
Artificial and Human Intelligence Integration
Every change the AI proposes travels through a review surface before it touches the work. The proposal card lists what the AI wants to do, rename a chapter, draw a connection, edit a passage. The writer reads each item, edits any input, unchecks anything they do not want, and approves the rest. The applied changes become a single step in the undo history, so an entire batch can be reversed with one keystroke. The AI never quietly mutates the work; the writer is always the editor. The intelligence runs on a key you bring yourself, Claude or Gemini, held in your own browser and never kept by Minerva, so the model answers to you and no one else.
Four commitments
Four commitments translate the philosophy into product decisions:
Many registers, not only text.
Thought shows up as diagrams, sketches, gestures, audio, images, models—not just prose. Minerva starts text-heavy because that register is densest today, but the architecture is built to make other registers first-class over time.
Retrieve + propose; the user decides.
The AI works on the user's own material. It surfaces, cross-references, and suggests—then everything goes through review. No silent edits; every change is inspectable and vetoable.
Wisdom over productivity.
Speed is not the goal. The goal is clearer thinking: better perspective, stronger connections, and deeper understanding of the work.
Holistic understanding.
The intelligence reads the work as a whole—chapters, references, connections, hierarchy—so it can spot drift, repetition, gaps, and underused sources.
Minerva is not a general-purpose chatbot, a prompt-to-content generator, a Word/Docs clone, a whiteboard, or a default real-time collaboration tool—and above all not a productivity app. It is a spatial map of thought, an editor as a thinking surface, and an AI partner grounded in retrieval, reasoning, and user-controlled review.
How it works
The philosophy is load-bearing only if the architecture carries it. The following sections describe the parts of the system that do the carrying.
The User Flow
The diagram traces how a person moves through Minerva. A user arrives at the public homepage, where the About page sits one step away as an open invitation to understand the project before entering. Authentication opens the bounded interface, and inside it the Projects Collection presents the user's personal archive. Opening a project reveals its Field, and from the Field a chapter opens into the Editor; the two exchange focus directly, so the writer moves between the spatial view of the whole and the close work of a single chapter without losing their place. Appearance and Settings sit as panels reachable from within, never replacing the working surfaces. The nested boundaries in the diagram are themselves the argument: the editor sits inside a project, the project inside the interface, the interface inside the authenticated zone, each frame marking a level of context the work belongs to. Every surface is oriented toward the same question: what does this mean in the context of the whole?
The Architecture
The diagram reads in three bands. At the top are the surfaces the user touches: the Field, a spatial canvas rendered with React Flow; the Editor, a writing surface built on TipTap; and the lighter Projects list and landing pages. Everything is built on Next.js, React, and TypeScript, styled with Tailwind. Beneath the surfaces is the state the application holds while a project is open, organized around a single dominant store that carries the live project, with the reducer as its one gate: every change, whether the user drew it by hand or the AI proposed it, passes through the reducer, which consults the rule library before committing. Below that is the persistence layer, where everything passes through a single storage facade, is compressed, and lands in the browser's IndexedDB, with localStorage as a fallback. Two things cross the boundary of the application: the user's work, which leaves as a portable .mnrv file they own outright, and AI requests, which leave through one brokered door to the external providers Minerva rents but does not own, Anthropic's Claude and Google's Gemini, kept symmetric so the choice between them lives in the user's settings. Almost everything inside the boundary is open source; the rented intelligences are the only proprietary dependencies. The shape of the diagram is the credibility argument: a modern, conventional stack, a clean line between what Minerva owns and what it rents, and a visible slot where cloud sync will plug in.
The Data Structure
The diagram shows a small project as what it actually is: one graph the user shapes. A Minerva project is composed of three kinds of card. A chapter holds the work being authored, prose, and in future versions any other register. A reference holds supporting material, research, citations, sources the work draws from. A note holds the writer's marginalia, quick ideas, reminders, fragments not yet ready to be written into a chapter. All three share the same internal shape, a title, rich-text content, annotations, sketches, a position on the Field, and a single field, the card type, sets the role each one plays. Cards are connected by two kinds of edge. A branch connection indicates one card contains or leads to another, and forms the project's hierarchy. An inform connection indicates one card supplies context to another's reasoning, without sitting above it in the hierarchy. A small library of rules governs which connections are legal: a reference cannot branch-parent a chapter, because references inform chapters rather than containing them; a note cannot inform anything, because the AI should not treat fragmentary thought as authoritative context; no connection may form a cycle. The hierarchy itself is stored as a plain table of parent-to-child rows, which is what allows a chapter to belong to more than one parent at once, the same child simply appears in two rows. There is no separate tree and canvas; the sidebar and the Field are two views of the same rows. This is the deepest structural fact about Minerva: every project is the same handful of primitives arranged in an arrangement unique to its maker, and that whole graph is what the AI reads as its reasoning context.
Artificial Intelligence Integration
The diagram shows the loop a piece of work travels each time the intelligence touches it: a version of the content, surrounded by the signals the user has placed on it, passes through review and returns as a new version. The AI participates at two levels. At the Field level, it reads the project's structure, the cards, their types, the connections between them, and reasons about the work as a system. At the editor level, it reads the chapter's body alongside all of its annotations: highlights, margin notes, comments, and markup. The writer's own marks become the AI's instructions. The model is never given an undifferentiated dump of text; it always receives a structured context that tells it where it is in the work, what informs the current chapter, and what the writer has already said they want. The AI's output is always a proposal. The proposal is always reviewable. The review is always the writer's.
Workflow Scenarios
A few scenarios make the grammar concrete.
A novelist sequencing a book.
Each chapter card holds a chapter of prose; reference cards hold research, period sources, character dossiers. The Field shows the spine of the book and the connective tissue that runs between chapters, what foreshadows what, which arcs share which themes. A chapter that belongs to two arcs sits under both parents, and the AI's reasoning narrows accordingly.
User flow: Open the Field → survey the chapter sequence and arc connections → open a chapter → annotate with highlights and margin notes → ask the AI to analyze the chapter in context of its arc → review the proposal card → approve or adjust → return to the Field to see how the change ripples through the structure.
The horizon
Minerva is built to grow register by register. The architecture is general by design, the proposal-and-approve flow, the connection grammar, the multi-parent hierarchy stay constant, and each new version extends the kinds of material the system can hold and the kinds of input it can read. The current version already carries two registers: writing and freehand drawing. A sketch on the Field is legible to the AI as additional signal; an arrow drawn between two cards is a relationship the system understands. What follows extends the same logic outward.
The maker & the inception of the idea
I'm Enol Vallina, a designer and architect whose experience working across cultures and disciplines that blend craftmanship with technology has raised the questions this tool is designed and built to answer. These experiences have incubated the question of how the multifaceted complexity of creative thinking can be compatible, and enhanced, with computer-thinking. The work draws on parametric design, the discipline of treating a project as a system of parameters and relationships rather than a collection of artifacts, and on the conviction that it is precisely the intersection between that systemic rigor and the direct, unmediated quality of human input that makes a thinking tool worth building. Applied to an aid for thought, that intersection could produce something more carefully considered, more meaningfully shaped, and more truly unique than what the artificial intelligence moment has so far produced.
The product's direction is toward making the user progressively less dependent on any single mode of input. Each version adds a new channel, a sketch, a voice note, a gesture, an image, through which thought can enter the system without being translated into text first. The goal is not feature accumulation; it is the gradual removal of friction between how a person actually thinks and what the tool can understand. The more registers Minerva can read, the more faithfully it can serve as a companion to the full range of human creative expression.
