paper-spine

PaperSpine Orchestrator

Use this skill as the suite entrypoint. PaperSpine exposes one main skill: paper-spine. Each stage is executed by reading the corresponding playbook under references/.

Update detection: If the user asks to update, upgrade, or check for updates, read references/update.md and execute without starting the full writing workflow.

Command Routing

Trigger	Read
Default / full workflow	This file (Non-Negotiable Route below)
resume / continue / "continue from"	references/resume.md
update / "check for updates"	references/update.md
submission / "submission materials"	references/submission.md
audit	references/audit.md
translate	references/translate.md
humanize	references/humanize.md
respond / response letter / reviewer comments	references/respond.md

For fine-grained stage execution, the orchestrator reads the playbook for each stage from references/ rather than requiring a separate skill invocation:

Stage	Playbook
Intake / config	references/intake.md
Resume / checkpoint	references/resume.md
Research	references/research.md
Citation	references/citation.md
Rewrite	references/rewrite.md
Build	references/build.md
Humanize	references/humanize.md
LaTeX	references/latex.md
Translate	references/translate.md
Respond	references/respond.md
Audit	references/audit.md

Historical worker skill names (paper-spine-research, etc.) are legacy only and are not user entry points.

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.

Contribution-First, Reviewer-Aware Rules (V4)

These three rules sit above the motivation thread. Motivation remains required, but it supports the contribution rather than replacing it.

1. Contribution-First. The manuscript's highest-priority organizing unit is the confirmed contribution. Do not begin substantive writing until confirmed_contribution.md exists (what the paper adds, what problem/gap/ challenge makes it necessary, what evidence validates it, what claim boundary to respect, why a reviewer should find it publishable). Template + per-section checklists: references/contribution.md. Gate: contribution_check.py.
2. Results-as-Validation. Each major Results subsection must validate at least one contribution promise; metric-only units with no contribution mapping are a failure. Record this in results_validation.md. Template: references/results-validation.md. Gate: results_validation_check.py (journal / conference / competition scenes).
3. Reviewer-Aware. Before claiming submission-ready, create reviewer_audit.md (reviewer value map + objection register + editorial fit), populating the objection register from the three structured_review reviewer agents. Template: references/reviewer-audit.md. Gate: reviewer_audit_check.py.

The Stage 12 Final Audit hard gate runs all three checks; do not declare the work complete while any of them fails.

User-Facing Language

When the user writes in Chinese, ui_language=zh, output_language=zh, or translation_package=zh, all user-facing communication must be in Chinese throughout the whole run, not only in the final completion report. This includes intermediate progress updates, status bullets, tool-result summaries, blocked messages, gate re-run notes, final delivery tables, and error explanations.

Do not write English progress sentences such as "Chinese .docx generated", "Now writing the word report", "All stages passed", "Deliverables", or "PaperSpine Workflow Complete" in those Chinese-facing runs. Use Chinese status phrases instead, for example:

• 中文 Word 文档已生成：final_paper/paper.zh.docx。正在写入 Word 检查报告并重新运行关卡检查。
• PaperSpine 工作流已完成
• 全部阶段已通过
• 交付文件清单

Tool names, file paths, command names, and required English manuscript text may remain literal, but explanatory prose around them must be Chinese.

Required Configuration

Prefer reading paper_rewriting_output/paper_spine_config.json. If it is missing, read references/intake.md and collect configuration.

Required fields:

Field	Allowed Values
workflow	rewrite_existing, build_from_materials
scene	journal, conference, report_review, competition
tier	flash, pro
output_language	en, zh
target_name	free text
materials_dir	path or empty
draft_path	path or empty
user_motivation	free text or empty
official_urls	list
special_requirements	list
word_output	none, docx
translation_package	none, zh
reference_mode	local_first, specified_paths, web
reference_paths	list of local reference folders/files; default ["."]
citation_target_count	integer; default 20
humanize_tier	none, light, medium, heavy

Non-Negotiable Route

Resume-first rule: Before starting any workflow, read references/resume.md and run progress_check.py against paper_rewriting_output/. If earlier stages have produced valid artifacts, start from the first incomplete stage. After executing that stage, run its gate, run the full progress check again, and continue until final audit is complete. Do not restart from intake unless the user explicitly requests a clean run or the output directory is empty.

Anti-skip rule: Each stage is a gate. After completing a stage, run its gate check before moving to the next. If the gate fails, route back to that stage — do not skip, do not hand-write the missing artifact, do not patch downstream. The gate script is progress_check.py --gate <stage_name>. Never use bulk placeholder generators such as generate_artifacts.py, quick_generate.py, or mock_artifacts.py to replace real stage work.

Default Word output: Word (.docx) is a standard required artifact. Produce it in the LaTeX stage unless word_output is explicitly set to none in config. If pandoc is unavailable, write a BLOCKED/FAIL note in latex_report.md and do not declare completion.

---

Stage 1 — Intake / Configuration

If configuration is missing or incomplete, run the terminal wizard from the intake UI via the absolute installed launcher $env:USERPROFILE\.claude\skills\paper-spine\scripts\launch_paperspine_ui.ps1 (or the .codex equivalent, or $HOME/.codex/skills/paper-spine/scripts/launch_paperspine_ui.sh on bash); do not ask the user to hand-write JSON or answer a long plain chat checklist when a terminal is available. 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.

Gate:

python scripts/progress_check.py paper_rewriting_output --gate intake

If FAILED: return to intake. Config must exist and be valid before any research.

Stage 2 — Research

Read references/research.md and execute the research stage. Research must index local references per reference_mode and reference_paths before any web collection.

Required outputs: reference_materials/source_index.md, research_dossier.md, exemplar_learning_dossier.md, style_profile.md, sota_gap_map.md, motivation_options_after_research.md.

Gate:

python scripts/progress_check.py paper_rewriting_output --gate research

If FAILED: return to research. All six artifacts must exist before citation work begins.

Stage 3 — Citation Support Bank

Read references/citation.md and build citation_support_bank.md. Generate at least citation_target_count * 3 candidates (default: 60). ~80% recent. Include Source Channel for every row.

Gate:

python scripts/progress_check.py paper_rewriting_output --gate citation

If FAILED: return to citation. The bank must exist with sufficient candidates.

Stage 4 — Motivation Confirmation

Stop for user confirmation of the controlling motivation. Write confirmed_motivation.md only after the user chooses.

This stage is BLOCKED (not just pending) until the user confirms. Present motivation_options_after_research.md and wait. Do not auto-select.

Gate:

python scripts/progress_check.py paper_rewriting_output --gate motivation_confirmation

If FAILED/BLOCKED: stop and wait for user. Do not proceed with an unconfirmed motivation.

Stage 5 — Humanize (if applicable)

If humanize_tier is light, medium, or heavy, read references/humanize.md and apply tier-specific constraints.

Stage 6 — Writing / Drafting

If workflow is rewrite_existing, read references/rewrite.md. If workflow is build_from_materials, read references/build.md.

Both workflows must create section_blueprints.md and writing_rationale_matrix.md before drafting.

Gate:

python scripts/progress_check.py paper_rewriting_output --gate planning

If FAILED: return to the writing stage. Blueprints and rationale matrix are mandatory — a paper without a documented design is not a PaperSpine paper.

Stage 7 — Integrity Audit

Run integrity audit before LaTeX assembly:

python scripts/integrity_audit.py paper_rewriting_output --markdown --write

Review BLOCKER findings. Return to the relevant stage for any BLOCKED dimension. Do not proceed to LaTeX with unresolved BLOCKERs.

Gate:

python scripts/progress_check.py paper_rewriting_output --gate integrity_audit

If FAILED: run the audit, fix BLOCKERs, re-run until gate passes.

Stage 8 — LaTeX / PDF / Word

Read references/latex.md for LaTeX assembly, PDF compilation, and Word output. Word (.docx) is a standard required artifact. Produce and check it unless word_output is explicitly none.

Citation mechanism (hard rule): Every in-text citation must be a real \cite{key} linked to a bibliography entry — either \bibliographystyle{unsrt} (or plain/ieeetr) + \bibliography{references.bib}, or a thebibliography block whose \bibitem{key} entries are reached by \cite{key}. Never type the bracket number as literal text: a hand-typed [1] is inert and does not link to the reference list (in the PDF or the .docx). A numeric bibliographystyle/CSL still renders as [1] or [3,12,13], so the visible plain-numeric style is preserved. Do not use author-year citations and do not wrap numeric citations in extra parentheses such as ([15]). latex_guard.py fails literal-bracket citations with no \cite and out-of-sync numbering; it also checks the format.

Title (hard rule): main.tex must contain \title{...} and \maketitle. Word output must begin with the paper title — not Abstract, Keywords, or body text. latex_guard.py checks the TeX source; word_guard.py checks the .docx.

When output_language is zh, the paper is Chinese. Produce final_paper/paper.zh.docx as the primary Word output instead of paper.docx — the .zh.docx suffix marks the language, not a translation:

pandoc paper_rewriting_output/final_paper/main.tex -o paper_rewriting_output/final_paper/paper.zh.docx --from latex --to docx --resource-path=paper_rewriting_output/final_paper --number-sections --citeproc --bibliography=paper_rewriting_output/final_paper/references.bib
python scripts/word_guard.py paper_rewriting_output/final_paper/paper.zh.docx --language zh --fix-fonts

--citeproc --bibliography=... resolves \cite to linked [1] numbers in the .docx (matching the English command). Use it when citations come from a references.bib. If the paper instead carries a thebibliography block with \bibitem, drop --citeproc --bibliography=... — pandoc resolves \cite against \bibitem natively; either way the source must use \cite, never literal bracket text.

python scripts/word_guard.py paper_rewriting_output/final_paper/paper.zh.docx --language zh --markdown --output paper_rewriting_output/word_report.zh.md

If output_language is en, produce final_paper/paper.docx as the primary Word output. Do not produce or require final_paper/paper.zh.docx unless translation_package=zh is explicitly requested.

Gate (LaTeX):

python scripts/progress_check.py paper_rewriting_output --gate latex

Word guard (run only for the Word files requested by the config):

python scripts/word_guard.py paper_rewriting_output/final_paper/paper.docx --language en --fix-fonts
python scripts/word_guard.py paper_rewriting_output/final_paper/paper.docx --language en --markdown --output paper_rewriting_output/word_report.md

When output_language=zh, run:

python scripts/word_guard.py paper_rewriting_output/final_paper/paper.zh.docx --language zh --fix-fonts
python scripts/word_guard.py paper_rewriting_output/final_paper/paper.zh.docx --language zh --markdown --output paper_rewriting_output/word_report.zh.md

Gate (Word):

python scripts/progress_check.py paper_rewriting_output --gate word --require

Use --require so the gate checks Word even when config says none — the file should exist and be valid. If word_output is explicitly none, skip this gate.

Stage 9 — Submission Package (if requested)

If submission materials are requested, read references/submission.md, create submission_package/, and run:

python scripts/submission_check.py paper_rewriting_output/submission_package --fix-fonts --markdown --write

Stage 10 — Translation Package (if applicable)

If output_language is en and translation_package is zh, read references/translate.md and produce the complete translation_zh/ package.

translation_zh/ is the translation audit/intermediate package, NOT the final user-facing Chinese document. The final Chinese deliverable is a single Word file under final_paper/. After translation_zh/full_paper_translation.zh.md is complete, generate the final Chinese Word document:

pandoc paper_rewriting_output/translation_zh/full_paper_translation.zh.md -o paper_rewriting_output/final_paper/paper.zh.docx
python scripts/word_guard.py paper_rewriting_output/final_paper/paper.zh.docx --language zh --fix-fonts
python scripts/word_guard.py paper_rewriting_output/final_paper/paper.zh.docx --language zh --markdown --output paper_rewriting_output/word_report.zh.md

Run python scripts/translate_guard.py paper_rewriting_output --markdown --write and require PASS. If pandoc is unavailable, write BLOCKED/FAIL in the translation report — do not silently skip the final Chinese Word document.

Gate:

python scripts/progress_check.py paper_rewriting_output --gate translation --require

Stage 11 — Review Response (if requested)

If the user requests review response / revision response, read references/respond.md, create review_response/, and run python scripts/respond_check.py paper_rewriting_output/review_response --markdown --write.

Stage 12 — Final Audit & Completion Hard Gate

Read references/audit.md. Before declaring the workflow complete, all checks below must pass. If any command fails or reports missing/content issues, the workflow is not complete; return to the failing upstream stage.

python scripts/artifact_check.py paper_rewriting_output --markdown --write
python scripts/citation_bank_check.py paper_rewriting_output/citation_support_bank.md --markdown --write
python scripts/progress_check.py paper_rewriting_output --gate final_audit
python scripts/integrity_audit.py paper_rewriting_output --markdown --write
python scripts/progress_check.py paper_rewriting_output --markdown --write

When word_output is not explicitly none, also run:

python scripts/word_guard.py paper_rewriting_output/final_paper/paper.docx --language en --fix-fonts
python scripts/word_guard.py paper_rewriting_output/final_paper/paper.docx --language en --markdown --output paper_rewriting_output/word_report.md

When translation_package is zh or output_language is zh, also run:

python scripts/word_guard.py paper_rewriting_output/final_paper/paper.zh.docx --language zh --fix-fonts
python scripts/word_guard.py paper_rewriting_output/final_paper/paper.zh.docx --language zh --markdown --output paper_rewriting_output/word_report.zh.md

Completion may be declared only when artifact_check.py exits 0, final progress reports is_complete=true, no misplaced_artifacts are reported, integrity audit has no unresolved BLOCKER, and Word output is present and valid unless the user explicitly opted out with word_output=none. If pandoc is unavailable, write BLOCKED/FAIL in latex_report.md; never silently skip Word.

Hard gate rules:

• artifact_check.md Status: FAIL or BLOCKED → workflow NOT complete; return to upstream stage.
• citation_bank_check.md Status: FAIL → citation bank unqualified; return to citation stage.
• Nested paper_rewriting_output/ inside paper_rewriting_output/ → misplaced; move contents up.
• Sibling final_paper/ outside paper_rewriting_output/ → misplaced; remove sibling copy.
• writing_rationale_matrix.md too thin (<8 rows, generic cells) → artifact_check FAIL → NOT complete.

Chinese completion report rule: When the output language is Chinese (output_language=zh) or a Chinese translation package is requested (translation_package=zh), the final user-visible completion report must be written in Chinese. Use Chinese section titles and status labels. Prohibited: "PaperSpine Workflow Complete", "All stages passed", "Deliverables". Required minimum content: 工作流已完成、全部阶段已通过、交付文件清单. The final Chinese Word document is final_paper/paper.zh.docx; the translation_zh/ folder is an audit/intermediate package, not the final deliverable.

---

Loop Rule: If a gate fails, route back to that stage. Do not patch the final paper directly when the missing artifact should have been created earlier.

If a worker skill is unavailable, follow the reference playbook locally and produce the same artifacts.

Migration Note

See references/orchestrator-branch-map.md for stage ownership details.

Standard Artifacts

Write workflow artifacts under paper_rewriting_output/.

final_artifact_manifest.md must label each artifact with its source category:

• required — always produced
• pro-extra — produced only in pro tier (additional analysis depth)
• optional-translation — produced when translation package is requested
• optional-submission — produced when submission materials are requested
• optional-review-response — produced when review response workflow runs

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.

Final artifacts: latex_report.md, final_artifact_manifest.md, final_paper/main.tex, final_paper/paper.pdf (when TeX available), final_paper/paper.docx + word_report.md (standard; skip only if word_output is explicitly none), final_paper/paper.zh.docx + word_report.zh.md (when output_language=zh or translation_package=zh; translation_zh/ is the audit/intermediate package, NOT the final Chinese deliverable), submission_package/ (when requested).

Writing Rationale Matrix

writing_rationale_matrix.md must be created before final writing in both workflows. It is the execution plan, not a post-hoc summary:

Row ID	Manuscript Unit	Current/Planned Function	Motivation Link	Reference/SOTA Pattern Learned	Target Scene or Venue Norm	User Evidence or Citation Anchor	Planned Change	Final Text Check

The first data row must justify the whole-work framework in depth. Subsequent rows split the document into the smallest useful writing units. Each row must explain concrete anchors across multiple dimensions. A shallow matrix is a failure.

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, read references/intake.md to launch the intake UI. In Claude Code, /paperspine must launch the intake UI automatically.