wubing2023/paperspine

PaperSpine is a writing skill suite for two AI coding hosts, Codex and Claude Code, aimed at people writing academic or technical papers. The pitch is that it is not a polishing tool. Before the agent writes anything, it is asked to learn the target scene, for example a specific journal or conference style, and to study strong recent examples in the same field. Then every unit of the manuscript it plans or changes is recorded with a reason. The supported target scenes are journal papers, conference papers, course or technical reports and reviews, and competition papers. There are two equally weighted workflows. One is rewriting an existing manuscript without treating the task as cosmetic editing. The other is building a paper from a folder of raw materials such as notes, figures, PDFs, data summaries, partial drafts, and experiment descriptions. Output can be in English or Chinese, and English runs can also produce a translation package containing Chinese versions of the intermediate artifacts and the final Markdown. The repository is laid out so that dist/codex/paper-spine is a single bundled skill folder for Codex, while dist/claude/skills/ contains a flat suite of seven smaller skills for Claude Code (intake, research, rewrite, build, latex, audit, and an orchestrator), plus slash-command helpers under dist/claude/commands/. Shared scripts and references sit in src/. A Windows PowerShell installer (install.ps1) copies the right pieces into either ~/.codex/skills/ or ~/.claude/skills/, and there is also a Claude Code plugin manifest under .claude-plugin so users can install via /plugin marketplace add and /plugin install. A complete run produces an audit trail under paper_rewriting_output/ rather than just a final document. The artifacts include a research dossier, motivation candidates and a confirmed motivation, a source inventory, an evidence bank, a figure asset map, a claim register, a section blueprint, a rewrite matrix, a revision audit, and the final paper as main.tex with references.bib and optional .docx and .pdf outputs. The central artifact is writing_rationale_matrix.md, which has to explain unit by unit what each section does, how it serves the confirmed motivation, what was learned from examples, what evidence supports it, and what final checks should pass. Quality is enforced by Python scripts: artifact_check.py validates the artifact set, latex_guard.py checks the LaTeX file, and word_guard.py checks the Word output. The README also lists what PaperSpine tries to prevent, including sentence-by-sentence polishing with no logic change, treating different genres as the same, writing before confirming the motivation, and adding unsupported claims. License is MIT.