paper-explainer-summarize
Digest an academic paper into a structured overview and perspective reports covering background, method, experiments, discussion, and related work. Use for conference or journal papers and preprints. Use pdf-explainer-summarize for books, manuals, dissertations, or other long non-paper PDFs.
- License 许可
- MIT
- 最近更新
- 2026-10-27
- Stars
- ★ 1
- Forks
- 0
01 前言 Front Matter
- name
- paper-explainer-summarize
- description
- Digest an academic paper into a structured overview and perspective reports covering background, method, experiments, discussion, and related work. Use for conference or journal papers and preprints. Use pdf-explainer-summarize for books, manuals, dissertations, or other long non-paper PDFs.
- user-invocable
- true
02 资源文件 12
- 参考文档docs/ocr-pipeline.md9.2 KB
- 其他flake.lock569 B
- 其他flake.nix1.2 KB
- 参考文档references/finalize.md13.1 KB
- 参考文档references/overview-template.md3.1 KB
- 参考文档references/phase-1.md15.2 KB
- 参考文档references/runtime-and-ocr.md8.9 KB
- 脚本scripts/dblp_bibtex.sh1.0 KB
- 脚本scripts/dblp_lookup.sh1.7 KB
- 脚本scripts/mineru_ocr.sh4.7 KB
- 脚本scripts/mineru_to_paper_md.py21.3 KB
- 脚本scripts/preflight.sh2.1 KB
03 SKILL.md 正文
Paper Summary Pipeline
Digest an academic paper into reports/overview.md (a TL;DR + key figure, the Ochiai format's 6 questions capturing novelty, usefulness, and validation, and short 前提知識 / 原文の読み方 sections) plus per-perspective detail reports written by parallel perspective agents. Papers are short (8–30 pages), so the heavy 3-phase pipeline of pdf-explainer is unnecessary — Phase 1 runs inline in the orchestrator; only the perspective reports fan out:
Phase 1 (orchestrator, inline) Phase 2 (paper-detail, parallel) Finalize
read paper (local MinerU OCR) → reports/background.md → consistency & faithfulness
biblio metadata + source-structure.md reports/method.md sweep over the whole set
write spine.md (confirmed facts) reports/experiments.md (paper-explainer-consistency-
write reports/overview.md reports/discussion.md sweep) → targeted fixes
(spine handed to every perspective) reports/related-work.md (dblp)
The work dir's internal layout carries reports, [pNN] anchors, a factual
spine.md, and a separate source-structure.md authority for the paper's own
section topology. This separation lets shared site authoring restructure reports
without presenting report-created headings as source sections.
Content-workflow invocation
When this skill owns the paper report phase, apply
[[explainer-content-workflow-base]] with the paper profile. Require the exact
immutable run request and its source identity, final work directory, report
scope, output actions, and source options. Use those values instead of asking
again or relying on an earlier session. The shared coordinator has already
resolved the citation-slug directory during read-only preflight; write directly
there and do not rename it. A create action must not overwrite an Artifact,
and replace applies only to its exact target.
Return exact produced paths/digests, report-consistency checks, and any structured blocker. In direct standalone use, retain the interactive and provisional-directory procedures below.
<SKILL_DIR> below is this skill's own base directory; the bundled scripts live at <SKILL_DIR>/scripts/. Reference them by that skill-root-relative path — no absolute or plugin-root paths.
Before starting — read the runtime contract
Read references/runtime-and-ocr.md completely before running any command or creating the work directory. It contains the original Gotchas, Prerequisites, naming convention, and work-directory layout without abridgment.
Step 0 — Confirm scope once, up front
One interactive gate before any work in standalone use; do not re-prompt between phases. In a content-workflow invocation, the run request is the resolved gate, so do not ask again. (OCR needs no question — MinerU runs locally and was already verified in Prerequisites; mention it will run and that a first run downloads models and takes a while.) In standalone use, ask:
- Detail reports — default: all five (background / method / experiments / discussion / related-work). Accept a subset or "overview only".
- Collect the source PDF into the work dir at Finalize? — note the answer now so it is not asked again.
Phase 1 — Read the paper and write the overview (inline)
Read references/phase-1.md completely and follow it end-to-end. When it reaches the overview template, read references/overview-template.md completely before writing reports/overview.md.
Phase 2 — Perspective detail reports (parallel)
Write one perspective detail report per in-scope perspective (background / method / experiments / discussion / related-work) by applying the paper-explainer-paper-detail skill — it holds the per-perspective report templates and the strict bibliographic constraints. Run the perspectives in parallel:
- Under Claude Code, dispatch one
paper-explainer-paper-detailsubagent per in-scope perspective (multiple Agent calls in one message) so each report is written in an isolated context and they run concurrently. - Otherwise, apply the
paper-explainer-paper-detailskill inline, once per perspective, keeping each report's work self-contained.
Either way, provide each perspective run only the inputs below; the report structure and constraints come from the skill itself, not from the orchestrator:
- The perspective (
background/method/experiments/discussion/related-work) - The source — OCR ran: absolute path to
<WORK_DIR>/ocr/paper.md; otherwise: the PDF absolute path plus the perspective's page span (from the section map, ±1 page margin; when in doubt, the span may generously cover the whole body — papers are short) - The section map (compact, from Phase 1)
- The assigned figure/table files — from
<WORK_DIR>/ocr/figures/, the files whose page falls in this perspective's span. Read each file's page from<WORK_DIR>/ocr/figures.md(the filename encodes the paper number, not the page). Every one must be explained. Assign each figure to exactly one in-scope perspective so none is duplicated or dropped: by its page → the perspective whose span contains it; when a page is shared, route by kind (teaser/motivating-example figures → background; architecture/overview/method figures → method; result plots/tables → experiments). Hold any figure whose page is covered only by an out-of-scope perspective for the overview (handled at Finalize). - The output absolute path
<WORK_DIR>/reports/<perspective>.md - The report language
- The spine — the absolute path to
<WORK_DIR>/spine.md. Every perspective gets it and treats it as authoritative for the facts it covers (thesis + direction, future-direction conclusion, running-example map, headline numbers + scope, figure-verified facts), rather than independently re-deriving them from OCR prose and diverging. The overview and each detail report are written independently and never reconciled afterward, so these facts must be taken from the one shared source, not rediscovered per report. (Thepaper-explainer-paper-detailskill states how a perspective consumes the spine, including the "figure wins over a wrong spine row" safety net.) - For
related-workadditionally: the absolute path to<SKILL_DIR>/scripts/dblp_lookup.shand the References[pNN]span
Perspective → relevant sections, when resolving spans: background → abstract + introduction + motivation/background sections; method → approach/method + preliminaries; experiments → experiments/evaluation + result appendices; discussion → discussion/limitations/conclusion; related-work → related work + introduction + references. (background and related-work both draw on the introduction — background from its motivation/problem framing, related-work from its cited prior work.)
Context hygiene: each perspective writes its report to a file and returns only a path + one-line status. Do not fold a full detail report back into the orchestrator context. The whole-set read that Finalize needs (all reports at once, to catch cross-report contradictions) is done by the consistency sweep in its own isolated context — only its findings list returns to the orchestrator, so hygiene is preserved even though the whole set is read. The orchestrator's own Finalize reads stay narrow: the "次に読むべき論文" section of related-work.md (for overview item 6), and the specific loci the sweep flags (to apply targeted fixes). (When dispatching to subagents this is automatic; when applying the skill inline, keep the same discipline — do the sweep as a self-contained pass and carry back only its findings.)
Finalize
Read references/finalize.md completely and follow it end-to-end. That reference contains the original Finalize procedure and Success criteria without abridgment.
Bundled scripts
scripts/preflight.sh <command> [args…]— resolves the skill's runtime env once (poppler + a MinerU source), then execs the command inside it: PATH mode when poppler + (uv, ormineru+python3) are present, elsenix developwhen aflake.lockis bundled andnixis available, else an aggregated fail listing both setup options (nix / manual). Launchmineru_ocr.shthrough it (preflight.sh bash …/mineru_ocr.sh <pdf> <ocr-dir>). Does not choose the MinerU source itself — the worker does that PATH-first (installedmineru, elseuvx --from 'mineru[core]' mineru).scripts/mineru_ocr.sh <pdf> <ocr-dir>— reads the PDF locally with MinerU (-b pipeline, text mode for born-digital PDFs / OCR mode for scanned ones, chosen by probing the text layer) and materializesocr/paper.md,ocr/figures/(figures/tables named by paper numberfig-NN/table-NN), andocr/figures.md(metadata index: label / file / page / caption, plus a "⚠ Not extracted" section for regions MinerU located but failed to crop) viamineru_to_paper_md.py. Launch it throughscripts/preflight.sh(which supplies poppler + a MinerU source). MinerU is resolved PATH-first: an installedmineruis used as-is, elseuvx --from 'mineru[core]' mineruresolves it into uv's shared tool cache (no manual install; first run syncs the tool env and downloads models — slow). How the two scripts turn MinerU's layout skeleton + the PDF text layer intopaper.md(and the fidelity trade-offs) is documented indocs/ocr-pipeline.md.scripts/dblp_lookup.sh "<title> <first-author surname>"— queries the public dblp API and prints candidate records as JSON lines ({key, type, title, authors, venue, year, doi, url}). No key needed; empty output = no hit. Used by the orchestrator to buildpaper.bib(and to verify related-work when that report is out of scope), and by therelated-workperspective (pass it the absolute path).scripts/dblp_bibtex.sh "<dblp-key>"— fetches the canonical BibTeX for a dblp record key (thekeyfromdblp_lookup.sh), retrying dblp's transient 503s. Exits non-zero and prints nothing on failure so the caller falls back to a hand-built entry. No key needed.
04 同仓库更多 Skills
invalid-skill-frontmatter
暂无简介
pdf-explainer-summarize
Digest a book, manual, or other long non-paper PDF into a source-anchored Markdown overview. Use when the requested deliverable is only a report, digest, or summary. Use paper-explainer-summarize for academic papers, direct reading for short PDFs, and book-explainer for coordinated reports, audio, site, rebuild, or resume.
pdf-explainer-pdf-stitch
Internal Phase 2 worker invoked once by pdf-explainer-summarize after chunk extraction. Reconciles sections across chunk boundaries, deduplicates the material, and writes the canonical structural spine and outline.
pdf-explainer-pdf-extract
Internal Phase 1 worker invoked by pdf-explainer-summarize once per page chunk. Reads the chunk visually or from a faithful text layer and writes structured source material with a verbatim heading stream, not a finished report.
pdf-explainer-pdf-detail
Internal per-chapter worker invoked by book-explainer for PDF input. Re-read one resolved chapter span visually or from a faithful text layer and write a standalone detail report with [pNN] source anchors.
pdf-explainer-generate-site
Compatibility entry point for requesting a local PDF reading site from an existing work directory. Route through book-explainer so Artifact reconciliation and Planning still occur.