Paper Spine

PaperSpine Orchestrator

Use this skill as the suite entrypoint. It is the main orchestrator: it routes the user to UI, intake, research, citation, rewrite, build, LaTeX, translate, humanize, audit, and update branch skills.

Update detection: If the user asks to update, upgrade, check for updates, or manage the PaperSpine installation, route immediately to paper-spine-update without starting the writing workflow.

Operating Principle

PaperSpine is a research-writing workflow, not a prose patcher. Its job is to learn the target scene and strong examples first, force a user-confirmed motivation, design the paper row by row, and only then write or rebuild the manuscript.

Never fabricate data, metrics, p-values, datasets, citations, figures, or experimental claims. User materials are authoritative for this paper's results. External examples teach structure and rhetoric only.

Required Configuration

Prefer reading paper_rewriting_output/paper_spine_config.json. If it is missing, route to paper-spine-intake or ask the same fields directly.

Required fields:

Non-Negotiable Route

1. If configuration is missing or incomplete, run the terminal wizard from paper-spine-ui, then paper-spine-intake; do not ask the user to hand-write JSON or answer a long plain chat checklist when a terminal is available. The wizard is the supported Claude Code/Codex command-line UI. In Claude Code, /paperspine is the preferred entry: it launches the external intake window automatically when configuration is missing.

   Codex hard constraint (do not skip): when paper_rewriting_output/paper_spine_config.json is missing, the FIRST tool action after reading this skill MUST be to run the absolute launch_paperspine_ui.ps1 (resolve under ~/.codex/skills/... or ~/.claude/skills/...) with sandbox_permissions: require_escalated so the window can open. Do NOT inspect materials, infer defaults, create config files, or continue the workflow before the UI has launched or explicitly failed. Only fall back to numbered/chat intake if the launch returns an error.

2. Always create or verify source_map.md.

3. Always use paper-spine-research before choosing the final motivation. Research must first index local/default references according to reference_mode and reference_paths; web collection supplements this index but does not replace it.

4. Research must create reference_materials/, research_dossier.md, exemplar_learning_dossier.md, style_profile.md, sota_gap_map.md, and motivation_options_after_research.md.

5. Use paper-spine-citation to create citation_support_bank.md. This bank is separate from exemplar learning: it supports Introduction, Related Work, Discussion, background, limitation, and application claims. Generate at least citation_target_count * 3 candidates; default target is 20, so the default candidate pool is 60. About 80% should be recent, using current_year - 3 as the simple threshold.

6. Stop for user confirmation of the controlling motivation. Do not write or rewrite until confirmed_motivation.md records the user's chosen motivation. The final motivation should be concise and specific. Do not inflate one narrow contribution into a multi-claim motivation.

7. If workflow is rewrite_existing, use paper-spine-rewrite.

8. If workflow is build_from_materials, use paper-spine-build.

9. Before drafting, both workflows must create section_blueprints.md and writing_rationale_matrix.md. The matrix is the execution plan, not a post-hoc summary.

10. Run the integrity audit: python scripts/integrity_audit.py paper_rewriting_output --markdown --write. This produces integrity_audit.md — a teaching report where every finding includes root cause, fix action, downstream impact, and a teaching note. The report must show no BLOCKED findings before LaTeX compilation can proceed.

11. Use paper-spine-latex for final LaTeX structure, figure placement, citation safety, and compile-oriented cleanup.

12. Always produce final LaTeX source. Compile PDF when a TeX engine is available. Markdown alone is not a final PaperSpine output.

13. If word_output is docx, produce and check a Word version.

14. If output_language is en and translation_package is zh, use paper-spine-translate to produce the complete translation_zh/ package. Run python scripts/translate_guard.py paper_rewriting_output --markdown --write and require PASS. The translation package must cover every required intermediate and final artifact with row-by-row translation of large tabular files. Summaries are not acceptable.

15. Use paper-spine-audit before declaring the work complete.

If another skill is unavailable, follow the referenced workflow locally and produce the same artifacts.

Standard Artifacts

Write workflow artifacts under paper_rewriting_output/ unless the user asks otherwise.

Common required artifacts:

• paper_spine_config.json
• paper_spine_config.md
• source_map.md
• reference_materials/source_index.md
• research_dossier.md
• exemplar_learning_dossier.md
• style_profile.md
• sota_gap_map.md
• motivation_options_after_research.md
• citation_support_bank.md
• confirmed_motivation.md
• section_blueprints.md
• writing_rationale_matrix.md

Rewrite existing:

• original_logic_map.md
• evidence_bank.md
• rewrite_matrix.md
• logic_transfer_audit.md
• revised manuscript

Build from materials:

• source_inventory.md
• evidence_bank.md
• figure_asset_map.md
• claim_register.md
• manuscript draft as an intermediate artifact

Final artifacts:

• latex_report.md
• final_artifact_manifest.md
• final_paper/main.tex
• final_paper/paper.pdf when a TeX compiler is available
• final_paper/paper.docx and word_report.md when Word output is requested
• translation_zh/ when English output requests a Chinese translation package

Writing Rationale Matrix Requirement

writing_rationale_matrix.md must be created before final writing in both rewrite_existing and build_from_materials. It must be a Markdown table used as the execution plan:

The first data row must justify the whole-work framework, structure, or main throughline in depth: why this controlling structure is chosen, how SOTA/target examples informed it, how it follows the confirmed motivation, which user evidence anchors it, and how the final manuscript will be checked against it. Subsequent rows must follow the target document in order and split it into the smallest useful writing units: paragraph-level moves, paragraph groups, model steps, assumptions, result/claim units, review synthesis units, competition solution blocks, headings, captions, and other argument-bearing fragments.

This is flexible by scene. A journal paper may naturally use abstract, introduction, methods, results, and discussion units. A competition paper may use problem restatement, assumptions, model construction, solving process, validation, sensitivity, strengths/weaknesses, and recommendations. A report or review may use executive summary, background, taxonomy, comparison, synthesis, and recommendation units. Do not force all tasks into a fixed IMRaD template.

Each row must explain concrete anchors across multiple dimensions: it advances or narrows the confirmed motivation, transfers a structural pattern learned from SOTA/example work, matches a target-scene norm, uses a user-provided evidence item, creates a front/back echo, fixes an original logic failure, and/or constrains a claim to available evidence. For important rows, write enough reasoning that the user can learn why this writing move is better.

A shallow matrix is a failure. If most rows say only "improve clarity" or "polish wording", stop and redo the research/blueprint stage.

Branch Map

Read references/orchestrator-branch-map.md when the workflow needs to be debugged or when a branch output fails audit. The rule is simple: route back to the branch that owns the weak artifact instead of patching the final paper directly.

Command-Line UI

Claude Code and Codex do not guarantee a native graphical picker for skills. The supported UI is the bundled terminal wizard. When configuration is missing, run paper-spine-intake. In Claude Code, /paperspine must launch the intake UI automatically; do not ask the user to call a separate UI command. The launcher opens the bundled terminal TUI, which supports Up/Down for option values, Left/Right for fields, Enter for edit or confirm, and S to save. Claude Code does not currently provide third-party skills with an API for embedding a custom keyboard UI directly inside the chat input box, so the real terminal TUI is the supported interactive path. Use native structured questions only when the host exposes them reliably in the current session. Use chat fallback only when terminal execution is impossible.