Ever After
Ever After
Ever After
Ever After
AI Development
AI Development
AI Development
Overview
Overview
My son loves to read, and as someone who grew up on Goosebumps and Choose Your Own Adventure books, I'm glad to get to share this love of reading with him. Over the years I've never lost the pull of a story that bends when you push on it - the thing video games understood and most books never tried. In the summer of 2026 I built the place I wanted those stories to live. Ever After is a self-hosted platform for illustrated, branching fiction. Readers pick a path and the book remembers. Nine books are published today, holding thirty-eight possible endings and about seventy hours of reading between them, from a four-minute bedtime story to a twenty-three-hour epic. Every one of them is free, and you can start reading without an account at all. Built in seven weeks: 1,091 commits, 4 AI applications (including local LLM using LM Studio), 14 shared packages, and more than 3,000 automated tests.
My son loves to read, and as someone who grew up on Goosebumps and Choose Your Own Adventure books, I'm glad to get to share this love of reading with him. Over the years I've never lost the pull of a story that bends when you push on it - the thing video games understood and most books never tried. In the summer of 2026 I built the place I wanted those stories to live. Ever After is a self-hosted platform for illustrated, branching fiction. Readers pick a path and the book remembers. Nine books are published today, holding thirty-eight possible endings and about seventy hours of reading between them, from a four-minute bedtime story to a twenty-three-hour epic. Every one of them is free, and you can start reading without an account at all. Built in seven weeks: 1,091 commits, 4 AI applications (including local LLM using LM Studio), 14 shared packages, and more than 3,000 automated tests.
Tech Stack
MidJourney
AI
ChatGPT
AI
Claude Code
AI
Higgsfield
AI
Docker
Development
Tech Stack
MidJourney
AI
ChatGPT
AI
Claude Code
AI
Higgsfield
AI
Docker
Development
Tech Stack
MidJourney
AI
ChatGPT
AI
Claude Code
AI
Higgsfield
AI
Docker
Development
Tech Stack
MidJourney
AI
ChatGPT
AI
Claude Code
AI
Higgsfield
AI
Docker
Development
Started
Started
2026
Pick Your Story Today
Pick Your Story Today
The Project Process
The Project Process
Exploring the Challenge
I started with a simple question and found four hard ones underneath it.
A branching book is not a book with links. If a reader's choice is going to mean anything, the system has to remember — not just where they are, but what they picked up, who they trusted, and what they can no longer go back and undo. That is a state machine, not a table of contents. It also has to be deterministic: the same choices must always produce the same story, or the reader's history is fiction of a different kind.
Reading is not one experience. A dyslexic ten-year-old, a parent reading in a dark room at 11pm, and someone on a bus with motion sensitivity are three different products wearing the same interface. Most reading apps hand you a font slider and call it accessibility.
"Anyone can publish" is a safety problem, not a feature. The moment strangers can put stories in front of children, you inherit every question a publisher has: Who rates this? What happens when the rating is wrong? How does a nine-year-old read at all when the law says they can't hold an account?
And AI is in the room whether or not you name it. These books are illustrated by AI, and some are co-written with it. Pretending otherwise wasn't an option I was interested in.
None of these had a single right answer. They needed positions I could defend.
The Design Approach
Design the container first. Before any screen existed, I designed .ever — a signed, checksummed package format that carries a story's graph, its illustrations, its metadata, and its rights attestations in one file. Every book on the shelf, including the ones I wrote, goes through the same validator as a stranger's submission. Nothing gets a back door because I made it.
Make accessibility an axis, not a checkbox. The reader exposes eight typographic controls — size from 80% to 250%, line height, letter, word and paragraph spacing, column measure set in characters rather than pixels, plus OpenDyslexic and Lexend — over six independent presentation axes: light/dark, normal/high contrast, reduced transparency, reduced motion, typeface, and theme. Each one that can defer to the operating system does, and stays deferred until the reader makes a real choice. The spacing controls deliberately range past WCAG 2.2's targets, so a reader can actually reach a conformant setting instead of merely being tolerated at the edge of one. Contrast is enforced by tests, not by intention.
Tell the reader what they're walking into. Every catalogue entry leads with how many endings a book holds and how long a single path runs — 13 minutes or 13 hours — before it asks for a commitment. The shelf sorts by mood and by the time you have, not by a genre tree.
Put the safety decisions in the model. A creator sets their own age rating; staff can override it, and a database constraint refuses any override that loosens one — the stricter of the two always wins, enforced in types so a careless call site fails to compile. Children under 13 don't self-register at all; a guardian holds the account and can see the shelf, which answers the COPPA question rather than lowering a number until it goes away.
Disclose the machine. AI involvement is a first-class field in the package format. Where a book was co-written with a model, the byline says so, publicly, on the shelf.
The End Results
Ever After is live, public, and free.
Nine published books, spanning a four-minute harbour story for early readers to a twenty-three-hour mystery — 38 endings and roughly 70 hours of reading in total. Cover art and interior plates are AI-generated and disclosed as such.
A reader anyone can tune. Day and night registers designed separately rather than mechanically inverted, three optional themes, and a preference set that survives across devices once you have an account — or works fine if you never make one.
A working path from stranger to shelf. Upload a package, read every validation finding in full, and follow the review from a private creator desk. The rating model, the validator, the importer, and the moderation vocabulary are production code, not a roadmap slide.
Engineering that holds it up. 3 applications and 14 shared packages behind more than 3,000 automated tests, including a real-browser suite across Chromium and WebKit — because a reader running an OS-level dark mode, at 250% text, with motion reduced, is a runtime a unit test cannot see.
And the part that isn't a metric: my son can open it, pick a story at his reading level, and change how it ends.
The Project Process
Exploring the Challenge
I started with a simple question and found four hard ones underneath it.
A branching book is not a book with links. If a reader's choice is going to mean anything, the system has to remember — not just where they are, but what they picked up, who they trusted, and what they can no longer go back and undo. That is a state machine, not a table of contents. It also has to be deterministic: the same choices must always produce the same story, or the reader's history is fiction of a different kind.
Reading is not one experience. A dyslexic ten-year-old, a parent reading in a dark room at 11pm, and someone on a bus with motion sensitivity are three different products wearing the same interface. Most reading apps hand you a font slider and call it accessibility.
"Anyone can publish" is a safety problem, not a feature. The moment strangers can put stories in front of children, you inherit every question a publisher has: Who rates this? What happens when the rating is wrong? How does a nine-year-old read at all when the law says they can't hold an account?
And AI is in the room whether or not you name it. These books are illustrated by AI, and some are co-written with it. Pretending otherwise wasn't an option I was interested in.
None of these had a single right answer. They needed positions I could defend.
The Design Approach
Design the container first. Before any screen existed, I designed .ever — a signed, checksummed package format that carries a story's graph, its illustrations, its metadata, and its rights attestations in one file. Every book on the shelf, including the ones I wrote, goes through the same validator as a stranger's submission. Nothing gets a back door because I made it.
Make accessibility an axis, not a checkbox. The reader exposes eight typographic controls — size from 80% to 250%, line height, letter, word and paragraph spacing, column measure set in characters rather than pixels, plus OpenDyslexic and Lexend — over six independent presentation axes: light/dark, normal/high contrast, reduced transparency, reduced motion, typeface, and theme. Each one that can defer to the operating system does, and stays deferred until the reader makes a real choice. The spacing controls deliberately range past WCAG 2.2's targets, so a reader can actually reach a conformant setting instead of merely being tolerated at the edge of one. Contrast is enforced by tests, not by intention.
Tell the reader what they're walking into. Every catalogue entry leads with how many endings a book holds and how long a single path runs — 13 minutes or 13 hours — before it asks for a commitment. The shelf sorts by mood and by the time you have, not by a genre tree.
Put the safety decisions in the model. A creator sets their own age rating; staff can override it, and a database constraint refuses any override that loosens one — the stricter of the two always wins, enforced in types so a careless call site fails to compile. Children under 13 don't self-register at all; a guardian holds the account and can see the shelf, which answers the COPPA question rather than lowering a number until it goes away.
Disclose the machine. AI involvement is a first-class field in the package format. Where a book was co-written with a model, the byline says so, publicly, on the shelf.
The End Results
Ever After is live, public, and free.
Nine published books, spanning a four-minute harbour story for early readers to a twenty-three-hour mystery — 38 endings and roughly 70 hours of reading in total. Cover art and interior plates are AI-generated and disclosed as such.
A reader anyone can tune. Day and night registers designed separately rather than mechanically inverted, three optional themes, and a preference set that survives across devices once you have an account — or works fine if you never make one.
A working path from stranger to shelf. Upload a package, read every validation finding in full, and follow the review from a private creator desk. The rating model, the validator, the importer, and the moderation vocabulary are production code, not a roadmap slide.
Engineering that holds it up. 3 applications and 14 shared packages behind more than 3,000 automated tests, including a real-browser suite across Chromium and WebKit — because a reader running an OS-level dark mode, at 250% text, with motion reduced, is a runtime a unit test cannot see.
And the part that isn't a metric: my son can open it, pick a story at his reading level, and change how it ends.
The Project Process
Exploring the Challenge
I started with a simple question and found four hard ones underneath it.
A branching book is not a book with links. If a reader's choice is going to mean anything, the system has to remember — not just where they are, but what they picked up, who they trusted, and what they can no longer go back and undo. That is a state machine, not a table of contents. It also has to be deterministic: the same choices must always produce the same story, or the reader's history is fiction of a different kind.
Reading is not one experience. A dyslexic ten-year-old, a parent reading in a dark room at 11pm, and someone on a bus with motion sensitivity are three different products wearing the same interface. Most reading apps hand you a font slider and call it accessibility.
"Anyone can publish" is a safety problem, not a feature. The moment strangers can put stories in front of children, you inherit every question a publisher has: Who rates this? What happens when the rating is wrong? How does a nine-year-old read at all when the law says they can't hold an account?
And AI is in the room whether or not you name it. These books are illustrated by AI, and some are co-written with it. Pretending otherwise wasn't an option I was interested in.
None of these had a single right answer. They needed positions I could defend.
The Design Approach
Design the container first. Before any screen existed, I designed .ever — a signed, checksummed package format that carries a story's graph, its illustrations, its metadata, and its rights attestations in one file. Every book on the shelf, including the ones I wrote, goes through the same validator as a stranger's submission. Nothing gets a back door because I made it.
Make accessibility an axis, not a checkbox. The reader exposes eight typographic controls — size from 80% to 250%, line height, letter, word and paragraph spacing, column measure set in characters rather than pixels, plus OpenDyslexic and Lexend — over six independent presentation axes: light/dark, normal/high contrast, reduced transparency, reduced motion, typeface, and theme. Each one that can defer to the operating system does, and stays deferred until the reader makes a real choice. The spacing controls deliberately range past WCAG 2.2's targets, so a reader can actually reach a conformant setting instead of merely being tolerated at the edge of one. Contrast is enforced by tests, not by intention.
Tell the reader what they're walking into. Every catalogue entry leads with how many endings a book holds and how long a single path runs — 13 minutes or 13 hours — before it asks for a commitment. The shelf sorts by mood and by the time you have, not by a genre tree.
Put the safety decisions in the model. A creator sets their own age rating; staff can override it, and a database constraint refuses any override that loosens one — the stricter of the two always wins, enforced in types so a careless call site fails to compile. Children under 13 don't self-register at all; a guardian holds the account and can see the shelf, which answers the COPPA question rather than lowering a number until it goes away.
Disclose the machine. AI involvement is a first-class field in the package format. Where a book was co-written with a model, the byline says so, publicly, on the shelf.
The End Results
Ever After is live, public, and free.
Nine published books, spanning a four-minute harbour story for early readers to a twenty-three-hour mystery — 38 endings and roughly 70 hours of reading in total. Cover art and interior plates are AI-generated and disclosed as such.
A reader anyone can tune. Day and night registers designed separately rather than mechanically inverted, three optional themes, and a preference set that survives across devices once you have an account — or works fine if you never make one.
A working path from stranger to shelf. Upload a package, read every validation finding in full, and follow the review from a private creator desk. The rating model, the validator, the importer, and the moderation vocabulary are production code, not a roadmap slide.
Engineering that holds it up. 3 applications and 14 shared packages behind more than 3,000 automated tests, including a real-browser suite across Chromium and WebKit — because a reader running an OS-level dark mode, at 250% text, with motion reduced, is a runtime a unit test cannot see.
And the part that isn't a metric: my son can open it, pick a story at his reading level, and change how it ends.
A human note on the tools
AI is not a replacement for real people — it's at its best amplifying human creativity, not standing in for it. It should not become a quiet excuse for organizations to cut staff purely to save money; the goal should be helping people do more, not erasing the people. And it isn't free: the energy and water that AI and its data centers consume carry a real environmental cost we shouldn't look away from. I built this with AI, eyes open — grateful for the leverage, and unwilling to pretend it's magic without consequences. The books published under the name D. S. Artden are my fictional stories that I crafted. The character arch, the emotional depth, the feeling of loss and love all come from me for these digital worlds I'm trying to create. AI was used to connect the dots, ensure my characters ending matches up over the course of the novel.
A human note on the tools
AI is not a replacement for real people — it's at its best amplifying human creativity, not standing in for it. It should not become a quiet excuse for organizations to cut staff purely to save money; the goal should be helping people do more, not erasing the people. And it isn't free: the energy and water that AI and its data centers consume carry a real environmental cost we shouldn't look away from. I built this with AI, eyes open — grateful for the leverage, and unwilling to pretend it's magic without consequences. The books published under the name D. S. Artden are my fictional stories that I crafted. The character arch, the emotional depth, the feeling of loss and love all come from me for these digital worlds I'm trying to create. AI was used to connect the dots, ensure my characters ending matches up over the course of the novel.
The Project Process
The Project Process
Exploring the Challenge
I started with a simple question and found four hard ones underneath it.
A branching book is not a book with links. If a reader's choice is going to mean anything, the system has to remember — not just where they are, but what they picked up, who they trusted, and what they can no longer go back and undo. That is a state machine, not a table of contents. It also has to be deterministic: the same choices must always produce the same story, or the reader's history is fiction of a different kind.
Reading is not one experience. A dyslexic ten-year-old, a parent reading in a dark room at 11pm, and someone on a bus with motion sensitivity are three different products wearing the same interface. Most reading apps hand you a font slider and call it accessibility.
"Anyone can publish" is a safety problem, not a feature. The moment strangers can put stories in front of children, you inherit every question a publisher has: Who rates this? What happens when the rating is wrong? How does a nine-year-old read at all when the law says they can't hold an account?
And AI is in the room whether or not you name it. These books are illustrated by AI, and some are co-written with it. Pretending otherwise wasn't an option I was interested in.
None of these had a single right answer. They needed positions I could defend.
The Design Approach
Design the container first. Before any screen existed, I designed .ever — a signed, checksummed package format that carries a story's graph, its illustrations, its metadata, and its rights attestations in one file. Every book on the shelf, including the ones I wrote, goes through the same validator as a stranger's submission. Nothing gets a back door because I made it.
Make accessibility an axis, not a checkbox. The reader exposes eight typographic controls — size from 80% to 250%, line height, letter, word and paragraph spacing, column measure set in characters rather than pixels, plus OpenDyslexic and Lexend — over six independent presentation axes: light/dark, normal/high contrast, reduced transparency, reduced motion, typeface, and theme. Each one that can defer to the operating system does, and stays deferred until the reader makes a real choice. The spacing controls deliberately range past WCAG 2.2's targets, so a reader can actually reach a conformant setting instead of merely being tolerated at the edge of one. Contrast is enforced by tests, not by intention.
Tell the reader what they're walking into. Every catalogue entry leads with how many endings a book holds and how long a single path runs — 13 minutes or 13 hours — before it asks for a commitment. The shelf sorts by mood and by the time you have, not by a genre tree.
Put the safety decisions in the model. A creator sets their own age rating; staff can override it, and a database constraint refuses any override that loosens one — the stricter of the two always wins, enforced in types so a careless call site fails to compile. Children under 13 don't self-register at all; a guardian holds the account and can see the shelf, which answers the COPPA question rather than lowering a number until it goes away.
Disclose the machine. AI involvement is a first-class field in the package format. Where a book was co-written with a model, the byline says so, publicly, on the shelf.
The End Results
Ever After is live, public, and free.
Nine published books, spanning a four-minute harbour story for early readers to a twenty-three-hour mystery — 38 endings and roughly 70 hours of reading in total. Cover art and interior plates are AI-generated and disclosed as such.
A reader anyone can tune. Day and night registers designed separately rather than mechanically inverted, three optional themes, and a preference set that survives across devices once you have an account — or works fine if you never make one.
A working path from stranger to shelf. Upload a package, read every validation finding in full, and follow the review from a private creator desk. The rating model, the validator, the importer, and the moderation vocabulary are production code, not a roadmap slide.
Engineering that holds it up. 3 applications and 14 shared packages behind more than 3,000 automated tests, including a real-browser suite across Chromium and WebKit — because a reader running an OS-level dark mode, at 250% text, with motion reduced, is a runtime a unit test cannot see.
And the part that isn't a metric: my son can open it, pick a story at his reading level, and change how it ends.
A human note on the tools
AI is not a replacement for real people — it's at its best amplifying human creativity, not standing in for it. It should not become a quiet excuse for organizations to cut staff purely to save money; the goal should be helping people do more, not erasing the people. And it isn't free: the energy and water that AI and its data centers consume carry a real environmental cost we shouldn't look away from. I built this with AI, eyes open — grateful for the leverage, and unwilling to pretend it's magic without consequences. The books published under the name D. S. Artden are my fictional stories that I crafted. The character arch, the emotional depth, the feeling of loss and love all come from me for these digital worlds I'm trying to create. AI was used to connect the dots, ensure my characters ending matches up over the course of the novel.