brycewang-stanford/paper-pipeline

Paper Pipeline

Overview

This skill runs the full post-first-draft polishing process on an academic paper by invoking five existing skills in a fixed order. It is an orchestrator: it does not duplicate any sub-skill's logic. Each stage is executed by invoking the corresponding skill with the Skill tool and following that skill's own workflow to full completion before moving on.

The pipeline (fixed order — never reorder or skip without explicit user request):

| Stage | Skill | Purpose in the pipeline | |-------|-------|------------------------| | 1 | paper-polish | Clean mechanical, consistency, and citation errors so the self-review can focus on substance | | 2 | paper-self-revise | Deep internal review and content revision on a clean draft | | 3 | paper-style | Restructure title + sections to the target journal, after content has stabilized | | 4 | paper-polish | Second pass: catch issues introduced by Stages 2–3 (cross-references, merge seams, new prose) | | 5 | reference-verify | Final reference audit, once the citation set is settled by all prior edits |

The order matters: polishing before reviewing keeps the review substantive; restructuring after content revision avoids churn; the second polish exists specifically because Stages 2–3 modify text and structure; reference verification runs last because earlier verification would be partially invalidated by later edits.

Input: A project folder containing main.tex (plus tables/figures, ref.bib). Output: The fully polished main.tex + ref.bib, per-stage backups, a ref_verify_report.xlsx, and a final pipeline report.

---

Phase 0: Setup (do all of this BEFORE invoking any sub-skill)

1. Receive folder path from $ARGUMENTS or ask. Verify main.tex exists with Glob.

2. Dropbox conflict check: Glob the project folder for *冲突副本* and *conflicted copy*. If found, stop and ask the user to resolve which version is canonical before anything else. Re-run this check between every stage (concurrent Overleaf editing can silently revert edits, and a 5-stage pipeline is a long exposure window).

3. Check for a previous pipeline run: if pipeline_state.json exists in the folder and has incomplete stages, show its status and ask: resume from the first incomplete stage, or start over from Stage 1? If resuming, skip directly to that stage.

4. Collect the target journal now (needed by Stage 3, asked upfront so the pipeline never stalls mid-run). If given in $ARGUMENTS, use it. Otherwise use AskUserQuestion with common options (JFE / JF / RFS / Management Science) plus "Other" for free-form input.

5. Ask for the interaction mode with AskUserQuestion — this is the single most important setup question, because five interactive skills back-to-back can mean hundreds of approval prompts:

   • 全程交互 — every sub-skill runs in its native item-by-item approval mode. Choose for final pre-submission runs.
   • 阶段确认 (recommended default) — within each stage, changes are applied without per-item prompts, but the pipeline pauses at each stage boundary to show that stage's summary and get a go-ahead before the next stage starts.
   • 全自动 — run all five stages straight through; only the final report at the end. The user's upfront choice of this mode counts as the explicit opt-in that each sub-skill's fast path requires.

6. Create the backup baseline: make a pipeline_backups/ subfolder and copy main.tex → pipeline_backups/main_stage0.tex and ref.bib → pipeline_backups/ref_stage0.bib. After each completed stage, snapshot main.tex as main_after_stage<N>.tex. These snapshots are the rollback path if a later stage goes wrong.

7. Write pipeline_state.json in the project folder:

{
  "folder": "<path>",
  "target_journal": "<journal>",
  "mode": "stage-confirm",
  "started": "<date via Bash>",
  "stages": {
    "1_paper_polish": "pending",
    "2_paper_self_revise": "pending",
    "3_paper_style": "pending",
    "4_paper_polish_2nd": "pending",
    "5_reference_verify": "pending"
  }
}

Update the relevant stage to "in_progress" when it starts and "done" when it completes. This file is what makes the pipeline resumable after an interrupted session.

---

Stage Execution Protocol

For every stage, follow this exact sequence:

1. Announce the stage with a banner so the user always knows where the pipeline is:

════════════════════════════════════════
  Stage N/5: <skill-name>  —  <one-line purpose>
════════════════════════════════════════

2. Mark the stage in_progress in pipeline_state.json.

3. Invoke the sub-skill with the Skill tool (e.g., skill: paper-polish, args: <folder path>). Then follow the loaded skill's instructions to full completion — all of its checks/comments/phases, including its own completion summary. Never paraphrase or shortcut a sub-skill's workflow; the whole point of the pipeline is that each stage runs exactly as it would standalone.

4. Apply the interaction mode:

   • 全程交互: follow the sub-skill's native approval flow unchanged.
   • 阶段确认 / 全自动: the user's mode choice in Phase 0 is the explicit authorization each sub-skill's fast path requires — treat it as the user having already said "accept all remaining" (paper-polish), "yes to all" (paper-self-revise), "直接改" (paper-style), and apply: auto (reference-verify). Still display each change block as the sub-skill specifies, so the user can scroll back and audit; just don't wait for input.

5. On stage completion: snapshot main.tex to pipeline_backups/, mark the stage "done", re-run the Dropbox conflict check, and record a 3–5 line stage summary (counts of changes applied/skipped, sections touched, anything flagged).

6. Stage boundary: in 阶段确认 mode, show the stage summary and ask "继续下一阶段？(yes / pause)". In 全自动 mode, continue directly. In 全程交互 mode, a brief "Stage N 完成，进入 Stage N+1" notice suffices since the user has been involved throughout.

---

Stage-Specific Notes

Stage 1 — paper-polish (first pass): Run all 19 checks as the skill specifies. Pass the project folder path as args.

Stage 2 — paper-self-revise: The skill's Phase 0 asks whether a review report already exists. In the pipeline context, prefer a fresh review: if a review.pdf/review.md from an earlier round is present, ask the user whether to reuse it or regenerate (stale comments may already be addressed); in 全自动 mode, regenerate without asking. The generated review excludes anything requiring new empirical work, which matches the pipeline's text-only scope.

Stage 3 — paper-style: Pass both the folder path and the target journal collected in Phase 0 as args, so the skill's Phase 0 journal question is already answered. In 全自动 mode the skill's fast path picks the title itself; when in doubt there, prefer keeping the current title (case-adjusted) — the title is the most subjective change in the whole pipeline and the easiest to get wrong silently.

Stage 4 — paper-polish (second pass): Run all 19 checks again in full. Expect the yield to concentrate in: hardcoded section references broken by Stage 3 renumbering, merge-seam transitions, abbreviations re-defined or orphaned by moved text, text–table consistency in paragraphs Stage 2 rewrote, and em-dashes introduced by new prose. A second full pass is deliberate — do not skip checks just because Stage 1 ran them.

Stage 5 — reference-verify: Run on the project directory (Excel output, default settings). Map the pipeline mode onto APPLY_FIXES: 全程交互 → interactive; 阶段确认/全自动 → auto. The skill makes its own timestamped backups before editing and writes ref_verify_changes.md.

---

Final Report

After Stage 5, mark all stages done in pipeline_state.json and present:

# Paper Pipeline 完成报告

论文: <title>    目标期刊: <journal>    模式: <mode>

| Stage | Skill | 应用 | 跳过 | 要点 |
|-------|-------|------|------|------|
| 1 | paper-polish | 12 | 3 | 修正 5 处文表不一致 |
| 2 | paper-self-revise | 9 | 1 | 审稿报告 14 条意见，强化引言张力 |
| 3 | paper-style | — | — | 10→8 节，标题改为 sentence case |
| 4 | paper-polish (2nd) | 6 | 0 | 修复 3 处硬编码 Section 引用 |
| 5 | reference-verify | 4 | 1 | 50 条引用，2 条元数据修正，报告见 xlsx |

备份: pipeline_backups/ （main_stage0 … main_after_stage5）
报告文件: review.pdf · ref_verify_report.xlsx · ref_verify_changes.md
遗留事项: <anything any stage flagged but did not fix, e.g. abstract over word limit>

The 遗留事项 list matters: sub-skills (especially paper-style Phase 5 and reference-verify) flag out-of-scope issues they deliberately don't fix — collect them all in one place so nothing silently evaporates.

---

Rules

• One stage at a time, in order. Never start a stage before the previous one's skill workflow has fully completed. Never run two sub-skills' steps interleaved.
• The pipeline is text-only. No stage re-runs regressions or changes empirical results; if a sub-skill surfaces an issue that would require new empirics, flag it in 遗留事项 instead.
• Partial pipelines on request: if the user asks to start from a specific stage or skip one ("跳过 paper-style"), honor it and record the skipped stage as "skipped" in the state file.
• If a stage fails or the user aborts mid-stage: leave the state file showing in_progress for that stage, summarize what was applied so far, and remind the user the pipeline can resume later by re-invoking paper-pipeline on the same folder.
• Communicate with the user in Chinese; all LaTeX content stays in English. No em-dash (---) in any newly written prose, per house rules.