How ShelfClub tags books
A book on ShelfClub can carry tags: moods, vibes, a pace, genres and content warnings. This page says what the tags are, where they come from, what a reader's vote does, how much of the catalog has them and how we are measuring them.
What the tags are
There are five kinds of tag, and each one is picked from a fixed list: 8 moods, 10 vibes, 3 paces, 19 genres and 28 types of content warning. A content warning also says how much of it there is: mild, moderate or severe.
Here is every tag, with what it is meant to mean.
Moods
- Cozy
- Low-jeopardy warmth: small stakes, a safe community, danger kept mostly off the page. A village mystery or a small-town romance qualifies; a tense thriller in a charming setting does not.
- Adventurous
- A journey, quest, voyage or exploration drives the book; physical action and new places are central.
- Emotional
- Grief, loss or hardship rendered so that moving the reader is the point of the book. Not for a book that merely has feelings in it.
- Dark
- Bleak, disturbing or morally grim; cruelty, despair or violence are dwelt on, not passed over.
- Hopeful
- Moves toward light: recovery, resilience, repair, things getting better by the end.
- Romantic
- A romance between characters is a central thread, whether or not sex appears on the page.
- Funny
- Written for laughs; comedy, satire or wit is a primary register.
- Spicy
- Explicit sex is depicted on the page, in detail. Closed-door, fade-to-black or tension without depiction is romantic, not spicy.
Vibes
- Beach Read
- Light, propulsive holiday reading.
- Page Turner
- Hard to put down: short chapters, constant hooks.
- Literary
- Prose style and form are themselves a point of the book.
- Thought-provoking
- Poses ideas or questions the reader carries away.
- Atmospheric
- A strong sense of place or mood; the setting is immersive.
- Character-driven
- Interior lives and relationships propel the story: the book's engine is people.
- Plot-driven
- Events, twists and momentum propel the story: the book's engine is what happens.
- Slow Burn
- Tension, romantic or otherwise, builds gradually.
- Epic
- Sweeping scope: a large cast, high stakes, a saga or several generations.
- Comfort Read
- A gentle, reassuring, low-stakes reading experience. Cozy is about the story's stakes; comfort-read is about how the book feels to read.
Pace
- Slow-paced
- Unhurried: the book lingers on description, interior thought or daily detail, and plot events are far apart.
- Medium-paced
- Neither rushed nor lingering: the story moves steadily, with room for scene-setting and reflection.
- Fast-paced
- Events come quickly: short scenes, frequent turns, little time spent between them.
Genres
Genres have no written definitions yet. These are their names.
- Science Fiction
- Fantasy
- Romance
- Mystery
- Thriller
- Horror
- Historical Fiction
- Literary Fiction
- Young Adult
- Children's
- Graphic Novels
- Poetry
- Biography & Memoir
- History
- Science & Nature
- Self-Help
- Business
- True Crime
- Religion & Spirituality
Content warnings
Some definitions name another type by its short name, which is shown beside each type.
General
- Violence
violence - Physical violence between people, shown or described: fights, assault, murder, war, gore. Sexual violence and torture have their own types.
- Sexual Content
sexual - Sexual content that is present but not explicit. Explicit, detailed scenes are graphic_sex.
- Strong Language
language - Strong or frequent profanity.
- Substance Use
substance - Drug or alcohol use, abuse or addiction.
- Death
death - Death and dying: a character's death, a terminal illness, or grief over a death as a major element.
- Mental Health
mental_health - Mental illness shown or discussed: depression, anxiety, PTSD, trauma, eating disorders. Suicide and self-injury are self_harm.
- Abuse
abuse - Physical, emotional or verbal abuse by someone with power or trust over the victim: domestic abuse, child abuse, neglect. Sexual abuse is sexual_violence.
- Self Harm
self_harm - Self-injury, suicide, a suicide attempt or suicidal thoughts.
- Animal Cruelty
animal_cruelty - Harm to an animal: abuse, cruelty, or an animal's death.
- Racism
racism - Racism or xenophobia: slurs, discrimination, or violence against people for their race or origin.
- Homophobia
homophobia - Prejudice, slurs or violence against LGBTQ+ people, transphobia included.
- Other
other - A warning that fits none of the other types.
18+ and adult themes
- Adult Content (General)
adult_content - Mature material for adult readers that no more specific type on this list covers.
- Sexual Violence
sexual_violence - Rape, sexual assault or sexual abuse, shown or recounted, including of a child.
- Dubious Consent
dubious_consent - Sex where consent is ambiguous, coerced, or given under a power imbalance.
- Kidnapping
kidnapping - A person is abducted or taken against their will.
- Stalking
stalking - Persistent unwanted following, watching or pursuit of a person.
- Captivity
captivity - A person is held or confined against their will: imprisonment, hostage-taking. Kidnapping is the taking; captivity is the holding.
- Trafficking
trafficking - Human trafficking: people bought, sold, or forced into work or sex.
- Graphic Sexual Content
graphic_sex - Explicit, detailed sex scenes.
- BDSM/Kink
bdsm - Bondage, dominance and submission, sadomasochism or other kink.
- Degradation
degradation - Humiliation or degradation during sexual acts.
- Breeding
breeding - Breeding kink or forced pregnancy. An ordinary pregnancy is not this.
- Incest
incest - Sexual or romantic relations between family members, step-family included.
- Manipulation
manipulation - Emotional manipulation, gaslighting or controlling behavior.
- Cheating
cheating - Infidelity: a partner in a committed relationship is unfaithful.
- Torture
torture - Deliberate, prolonged infliction of pain on someone who cannot escape it.
- Blood Play
blood_play - Blood used in a sexual or ritual context.
Severity
- Mild
- Briefly referenced.
- Moderate
- Significant scenes.
- Severe
- Graphic or pervasive.
These definitions are the wording the people labelling books work from. The model is prompted with wording of its own.
Where they come from
- A language model reads a book's title, authors, description and subject headings, and for some books other published information about the book, and picks tags from the fixed lists above. It is an open-weights model (Qwen3, 8 billion parameters) that runs on ShelfClub's own machine. No book text and no reader data is sent to an outside AI company for tagging.
- It only tags books it has enough to read about: a description of 200 characters or more that is not library boilerplate, or, for some books readers know, other published information about the book of that length. That is about a quarter of the catalog, and it has not reached all of those yet.
- Books it has not tagged may still carry tags from simple keyword rules over the description, and a pace guessed from page count. Those are rougher.
- Readers can vote for a mood, a vibe or a pace on a book's page. Enough of those votes put a tag on the book. The next section says how.
What a reader's vote does
Anyone signed in can vote for a mood, a vibe or a pace on a book's page, and take the vote back. A vote is kept as soon as it is cast. What it changes on the book depends on who else has voted.
- A mood or a vibe the book does not have is put on it once three readers of the book have voted for it. With fewer, the votes are kept and the book's tags stay as they were.
- A reader of the book is someone who has it on their shelf as reading or finished. Other people's votes are kept too, and start to count when the book is on their shelf that way.
- A book has one pace. It changes to the pace the most readers of the book voted for, once at least three have voted for that pace.
- Votes do not take away a mood or a vibe that the model or the keyword rules gave the book. They add to those, and a mood or a vibe that three readers voted for is listed first.
- A tag that readers put on a book comes off again when fewer than three readers of the book are behind it. That is worked out each time someone votes on the book.
Three is a small number, chosen so that one person, or two, cannot change a book's tags. A tag that three readers voted for is those readers' view of the book. The scoring described below is of the model's tags, not of votes.
How much of the catalog is tagged
These are counts of the books in the catalog, taken on . A book can be counted in more than one row.
| Books | Count | Share of the catalog |
|---|---|---|
| All books in the catalog | 1,714,735 | 100% |
| Tagged by the model | 476,389 | 27.8% |
| Checked for content warnings by the model | 462,773 | 27.0% |
| With at least one mood | 429,827 | 25.1% |
| With at least one vibe | 443,416 | 25.9% |
| With a pace | 1,602,883 | 93.5% |
| With at least one genre | 913,988 | 53.3% |
| With at least one content warning | 175,862 | 10.3% |
What a missing warning means
A book with no warnings listed has not necessarily been checked. The model has looked at a book only where it has tagged it. For every other book an empty list says nothing.
A warning list is a help, not a guarantee. It rests on the description more than on the text of the book, so it can miss what a publisher's description leaves out.
How we are measuring them
400 books, sampled across popular and obscure titles, short and long descriptions and every genre, are each being labelled by three people who do not see the model's answer first. The model will be scored against them:
- For content warnings: how many of the warnings any labeller named it also named.
- For the rest: how often its tags match the majority.
Each score will be shown next to how often the labellers agree with one another, since that is the ceiling.
The results will be published on this page, with their margins of error, including where they are bad. Until then no figure for the tags' correctness exists, and we do not claim one.
Corrections
We found that the keyword rules had been matching inside other words ("therapeutic" contains "rape", "heroine" contains "heroin", "question" contains "quest"). We fixed the rule and removed about 66,000 wrong tags from 61,000 books and 30,000 wrong warnings from their editions.
Tell us when one is wrong
Email [email protected] with the book and the tag.
On a book's page, "Is a tag or warning wrong?" is where a reader corrects either. A report of a tag or a content warning that does not fit comes to us: when we agree, we take it off the book, and neither the model nor readers' votes put it back. A content warning a reader adds there is on the book's page at once.