# Shelf Showdown

> A to-be-read pile picker. Paste the books you keep meaning to read with their back-cover blurbs, say what you are craving and which tropes are hard nos, and settle what to read next in a head-to-head bracket. https://shelf-showdown.skillsafe.ai/

## What it does

- Free, in the browser: parses a pile (a title line such as "Title by Author" followed by the blurb, one block per book, up to 16 books; or a reading-site CSV export with a Title column, filtered to the to-read shelf), draws it as a shelf of spines, prescans every blurb against a fixed list of 28 tropes and 5 content flags (cheating, love triangle, cliffhanger or series opener, animal harm, grief), benches books that match a hard no, and seeds a single-elimination bracket with byes. You tap the winner of each matchup, or fill the rest by seed (your own picks are kept).
- Metered task `tag` (blurb read): for every book, a verdict (match, maybe, hard_no, thin), tropes from the fixed list each with a quote copied from the blurb, heat (closed door to explicit) and darkness (light to very dark) estimates, pace, mood words, and how strongly the blurb promises each of five reader reward channels (a world to sink into, sentence pleasure, minds to model, easy momentum, need to know). The page re-seeds the bracket from the read: matches first, then maybes, then thin blurbs; hard nos are benched.
- Metered task `ref` (referee): any two books from the pile, judged on 2-5 criteria taken from the craving, with a second pass in swapped positions. Verdict left, right, toss_up (the passes disagree or it is even) or neither (neither delivers the craving). A left or right call can be sent straight into the bracket.
- Every reply is checked on the page: each trope quote must appear in that book's blurb, hard nos must not come back as matches, blurbs under 25 words must read as thin, heat and darkness estimates above your limits must not be matches, and the swap check must agree with the verdict.

## Honesty notes

- Heat and darkness are estimates read from the blurbs you paste, never from the books themselves. The model is told never to add plot, reviews, page counts or series facts the blurb does not state.
- Seeds are computed by the page, not by the model.
- The sample piles on the page are invented: every title, author and blurb was written for the page.

## Developer surface

- API tutorial: https://shelf-showdown.skillsafe.ai/api.html (base https://api.skillsafe.ai/v1/app-api; fields task, reader, books or left/right, prescan, tags, note - object fields JSON-encoded as strings)
- The parser, prescan, seeding and reconcile run in node too: https://shelf-showdown.skillsafe.ai/shelfkit.js

## Sources

- Built on the SkillSafe skills @haowjy/writing-principles (the reader reward channels) and @shipshitdev/advanced-evaluation (pairwise judging with position-swap bias mitigation).
- Model: gpt-terra via SkillSafe. Runs cost credits; the shelf, prescan, bracket and exports are free.