Ufuk Karaca

Spine is an editorial workspace that turns a novel into a navigable graph. Scenes are commits, plotlines are branches, characters are contributors with line-level blame. You drop in a manuscript and get back its structure - which threads weave together, which fizzle out, which characters drive which arcs. **The problem.** Developmental editors read hundred-thousand-word drafts to find structural issues: subplots that get dropped, character arcs that flatten, pacing that sags between climaxes. It's a job that lives in the editor's head - they read for weeks, take longhand notes, and slowly assemble a structural picture. The knowledge is hard to share, hard to teach, and impossible to query later. **What Spine does.** It uses Opus 4.7's 1M-token context window to hold the entire manuscript - and treats Claude as a literary editor, not a chatbot. The hero feature is the **agentic editorial brief**: click one button, Opus 4.7 reads all 321 scenes of *Pride and Prejudice* through 1M context, calls eight read-only tools to verify each claim against the literal scene text, and writes a real editorial brief - the kind a senior dev editor would hand back. *"A structurally exemplary three-act courtship novel. Volume one plants every bomb. Volume two detonates them at Hunsford. The looseness is in the post-crisis stretch where the Lydia rescue is processed in many small repetitive beats."* That's a real sample from a real run. Specific. Cited. Verifiable. **Other surfaces.** A tension ribbon along the top of the workspace shows every scene's importance over time, color-coded by plotline - the shape of the book at a glance. An "Ask the manuscript" surface uses the same agent loop with UI-control tools (`highlight_character`, `focus_plotline`, `open_scene`) so the model literally drives the workspace as it answers. A plot graph view surfaces orphan threads (never connect) and unresolved threads (trail off without merging). Characters can be talked to in their own voice, with full-book context. **Who benefits.** Developmental editors at agencies and publishing houses. Authors revising long-form drafts. MFA students. Literary scholars who want to query narrative structure quantitatively. The thing that took weeks now takes seconds - and the structural read is shareable, queryable, and persistent. **Stack.** Next.js 15, React 19, TypeScript strict, SQLite via better-sqlite3 with raw SQL. Three Claude tiers: Haiku 4.5 with prompt caching on the per-scene ingest hot path, Sonnet 4.6 for mid-weight reasoning, Opus 4.7 with adaptive thinking on the synthesis surfaces. Pioneer (Fastino's GLiNER2) for character disambiguation NER. Tavily for real-world research when Ask triages a question to it. Built end-to-end with Claude Code.