more-paper-workflow

Imported from bingyunjiang/more-paper-workflow (SKILL.md). Install upstream with npx skills add bingyunjiang/more-paper-workflow. Copyright stays with the author.

Skill metadata

version: v1.0.28-20260814 (2026-08-14) author: Dr. Jiang Bingyun（江博士） wechat: Bingyunjiang category: research license: CC BY-NC-SA 4.0 series: more series related_skills:

• more-comic-digitizer: "A sibling skill in more series for traceable, non-destructive digitization of children's hand-drawn comics."
• science-direct-cdp-pipeline: "Overlaps on CDP ScienceDirect download; this skill adds the full 8-step workflow from topic definition to paper polishing."
• zotero-review-matrix-skill: "Source of literature-review-matrix-schema, literature-review-docx-guide, and gbt7714-2015-citation-format references. Integrated into Step 7 writing preparation and writing."

Entry routing

• 每次调用本 skill 时，在选择 Step 路由前先执行 python3 "$SKILL_DIR/scripts/check_skill_update.py" --json；should_prompt=true 时按 references/update-reminder-protocol.md 软提醒用户，检查失败、网络不可达或 should_prompt=false 时直接继续当前任务。
• 对外只保留一个主入口 README.md / SKILL.md，用户只需要记一个入口。
• related_skills 和 README 中的 more 系列索引只用于选型与边界说明，不表示运行时依赖、自动组合或共享项目状态。
• 主名称统一为 more-paper-workflow；旧名称 more-paper-workflow-pro-skill 仅作为兼容触发别名保留。
• 对话式工作流支持 Step 1-8 直接进入，不把前序流程当成硬门槛。
• 快速导流先看 references/entry-guide.md，按任务意图选择下面的精确入口；不要把概括性说明推断成并不存在的统一短文件名：
  • Step 1: agents/step_1_entry.md → agents/step_1_topic.md
  • Step 2: agents/step_2_outline.md
  • Step 3: agents/step_3_entry.md → agents/step_3_search_plan.md
  • Step 4: agents/step_4_search_score.md
  • Step 5: agents/step_5_download.md
  • Step 6: agents/step_6_entry.md → agents/step_6_zotero.md
  • Step 7: agents/step_7_entry.md → agents/step_7_writing.md
  • Step 8: agents/step_8_entry.md → agents/step_8_polishing.md
• 参考文件的功能分组见 references/reference-index.md，新增或重排 reference 前先更新索引。
• Cross-agent entry vocabulary lives in references/entry-routing-index.md.
• Full trigger vocabulary lives in references/trigger-catalog.md.
• Scientific figure backend selection and strict reproduction live in references/scientific-figure-reproduction.md.
• Native Step 7 process, architecture, data-flow and sequence diagrams use references/paper-diagram-contract.md.
• 黑白刊发图使用该合同的 style=minimal：纯黑白、无底纹，并通过通栏字号门；inspect.svg 仅用于审阅，不得作为投稿图。
• Platform-specific launch hints live in thin adapter files such as agents/openai.yaml and .claude-plugin/marketplace.json.
• Step-specific runtime contracts use only the exact agents/ paths listed above plus the references routed by each entry file. If an exact file is absent, report an incomplete installation; do not invent a shortened filename.
• Raster/PDF numeric recovery follows inspect → spec-review → explicit spec confirmation → candidates.csv → quality assessment → review-decisions.json → observations.csv → formal data.csv → VisualSpec → render/QA. Candidate or observation files never enter VisualSpec.
• Step 7 figure work starts with scripts/build_figure_asset_check.py; MinerU candidates degrade through manifest.json -> full.md -> images/ -> PDF direct, and only selected original assets are materialized into the project.

Runtime rules

• Completion gates: references/completion-gates.md
• Failure triage: references/failure-triage.md
• Agent execution discipline: references/agent-execution-discipline.md
• Step handoff: references/step-handoff-contract.md
• Shared machine contract registry: schemas/workflow-contract-registry.json
• Shared run state: static/core/workflow-run-envelope.md / scripts/workflow_run_envelope.py
• Update reminders: references/update-reminder-protocol.md
• Direct-entry artifact graph: .skill-state/artifact_passport.json（Artifact Passport 只负责 Step 4-8 的材料识别、readiness 和非锁定路由，不覆盖 Step 1-3 的前期构思合同。）
• Artifact Passport keeps route_mode as non-locking route metadata for Step 4-8 direct-entry handoff.
• 质量门按当前 target_state 和实际能力分支启用：draft_ready、quick-polish、只读审计或 core 写作不得被投稿级、严格复现、CJK 图形、DOCX 导出或登录型下载环境门误阻塞。
• Require a dedicated runnable script for every reproduced figure, or a documented per-figure function when a batch runner is the stable entrypoint; record every path in manifest per_figure_scripts.
• Keep extraction_status, review_status, render_status, and delivery_status independent. A successful render never authorizes extraction/review, and an unreviewed candidate never implies formal-data or publication-ready delivery.
• For raster measurement, bind the exact input SHA-256 and original dimensions. Never measure a preview, resized copy, overlay, or recreation; enlargement is review-only.
• Native candidate digitization covers color-distinct lines, compact filled scatter markers, and vertical solid-color bars/histogram bins; all require explicit calibration and accepted original-resolution overlay review.
• For curves, evidence pixel gaps remain in observations.csv; only user-confirmed curve_topology controls formal breaks in style-independent data.csv. Any guide_constrained densification is written into data.csv with guide/residual/hash lineage, never generated by render.
• Official XLSX/CSV values may validate an immutable image-derived extraction but must not silently overwrite image-derived candidates, observations, or formal lineage.

Global discipline

• 张力先行：每一步先说清楚当前任务的核心拉扯。
• 双向校准：同时防止“不够”和“过头”。
• 最小对比：关键边界用最小对比例子校准。
• 反模式命名：把常见错误写入对应 Step 的检查项。
• 任务定义优先于实现定义：先判定问题、交付物和边界，再选动作。
• 读证据先于生成：先读取用户材料、检索结果、PDF/Zotero/已有稿件，再输出判断或正文。
• 假设显式化：任务边界、证据基础、数据源选择和写作范围必须先说明。
• 最小充分产物：只生成当前 Step 需要的最小可用产物，不强行扩展全流程。
• 保守修改既有稿件：润色和修订只处理明确范围，不新增未经确认的证据。
• 失败先分层：检索、下载、Zotero、写作失败先定位层级，再给最小补救动作。
• 完成声明过门槛：未满足 completion gate 时，不使用“已完成/没问题”等表达。
• 契约变更 vs 实现变更：改输入输出、证据等级、路由或完成标准时，视为契约变更。
• 快速通道不跳质量门：direct-entry 可以跳过前序流程，但不能跳过证据边界、登录门控、Zotero 写入确认、引用审计或 Step 8 保守修订边界。
• Checkpoint 是“当前 Step 的输入与风险确认协议”，不是线性流程锁。
• 不限制 Step 6/7 直接入口。
• 防截断原则：机器工件禁止截断；Markdown/XLSX/PDF 仅作为展示层可截断，必须保留稳定回查 ID，且不得反向污染 JSON、BibTeX、Zotero 映射或下载 manifest。
• 领域状态与全局状态分离：每个 Step 保留自己的 domain_state，跨 Step 统一映射为 readiness / can_continue / blocking / warnings / recommended_next_step；can_continue=true 不等于当前 Step 已完整完成。

Workflow map

• Step 1: research topic clarification and topic review
• Step 2: outline generation and outline optimization
• Step 3: search planning
• Step 4: literature search, scoring, and reporting
• Step 5: batch download routing
  • Stable artifacts: download_manifest.json, download_attempts.jsonl, pdf-附件池索引.json.
  • Locked execution discipline: preserve the existing source order and phase order; all download/CDP phases remain serial, and deprecated parallel flags remain ignored.
  • Manual recovery: after user-provided PDFs, run scripts/step5_reconcile_pdf_pool.py --output <dir> to reconcile without changing filenames.
• Step 6: Zotero organization and attachment consistency
• Step 7: writing, evidence matrix, style learning, original-figure-first insertion, native paper diagrams, explicit quick/reproduction figure routing, citation audit
• Step 8: conservative polishing and verification

Step 4 public sequence: 4.4 筛选依据 → 4.5 五维 → 4.6 T1-T4 → 4.7 引文扩展 → 4.8 饱和 → 4.9 报告生成与完成检查。

Step boundaries

• Step 1-6 负责定题、检索、下载和证据整理。
• Step 7 负责正文写作、证据矩阵、风格学习、图表/流程图生成、复现与引用审计；论文 PDF/MinerU 原图默认直接插入并使用 not_applicable。可信定量数据新图走 quick，论文流程图/架构图走 diagram，明确要求重绘或数字化时才进入 source-locked figure evidence pipeline / reproduction。
• Step 8 负责成稿级精修、终验和保守修订。
• 本 skill 自己拥有论文项目的 Step 1-8 合同、证据边界和完成门；不要把其他 more-* skill 作为前置步骤或自动后续步骤。
• README.md 只保留对外简洁入口，不承载运行态真相。
• 入口收敛不等于流程拦截；对话式工作流的 Step 1-8 都可以按 direct-entry 合同进入，Artifact Passport 的材料/readiness 路由范围保持 Step 4-8。
• 推荐 Zotero 用户安装 llm-for-zotero 插件。
• evidence_pack 是无 Zotero/MinerU 时的本地证据包入口。
• full-document / review-only / abstract-only / chapter-only / continue-existing / revision-only 仍是 Step 7 的公开模式名。

Public-first entry examples

README 首屏只保留 public-first entry examples；运行时细则仍以本文件和上方列出的精确 agents/ 路径为准。

• Step 1: use the topic-clarification entry prompt in README.md
• Step 5: use the direct download entry prompt in README.md
• Step 7: use the writing entry prompt in README.md

Update policy

• SKILL.md is the canonical version source.
• README.md and CHANGELOG.md are display copies and must stay in sync.
• scripts/check_skill_update.py compares the three version surfaces and reports any mismatch.
• Remote updates are detected by reading the current upstream branch's SKILL.md; a different commit with the same version does not trigger a version prompt.
• scripts/perform_skill_upgrade.py only updates a clean worktree with a configured upstream and fast-forward history; otherwise it leaves the current version untouched.