Release the kraken on your reading list.
English | 한국어
PaperKraken is a Claude Code skill that swallows a research-paper PDF and produces a tutorial-grade critical-reading report — a clean, typeset PDF that walks you through the paper the way an experienced researcher would read it: background first, every equation dissected, figures decoded element by element, and the authors' claims put on trial.
It is built on the 7-step critical-analysis methodology from Jacques Cornwell's Nature career column, "Seven steps for critically analysing research papers" (2026) — the 1–2 hours of expert reading, performed for you, with the process shown.
One command in, one PDF out. For the Transformer paper, that PDF is 25 pages (English) or 22 pages (Korean) containing:
| Section | What it does |
|---|---|
| 📖 How to read this report | Teaches Cornwell's 3-phase / 7-step method itself, so the report doubles as training |
| 🪜 Prerequisite ladder | Background concepts stacked in dependency order — each rung with a real-number worked example and "which equation this unlocks" |
| 🗺 Related-work map | Lineages of prior approaches → each one's limitation → where this paper fits, with web-verified citations |
| 🔬 Steps 1–3: Aerial view | Abstract dissected sentence by sentence; the core hypothesis quoted, translated, and unpacked; the knowledge gap mapped |
| 🧮 Step 4: Every equation, one card each | Exact LaTeX transcription → full symbol table → term-by-term functional roles → numeric intuition → dependency chain. Theory papers get theorem/proof cards instead |
| 🖼 Original figures, decoded | Figures are cropped from the source PDF (never redrawn) and read element by element: every box, arrow, and symbol |
| 📊 Step 5: Independent conclusions | Results tables re-typeset and cell-verified; "what the numbers alone say" strictly separated from "what they can't say" |
| The Top 3–5 places readers actually stall, each resolved: misreading → why the paper invites it → resolution | |
| ⚖️ Steps 6–7: The verdict | Author claims vs. independent reading, an internal-consistency audit (it caught the Transformer paper saying 41.0 BLEU in §6.1 vs 41.8 in Table 2), confounders, and conflicts of interest |
| An appendix of copy-paste numpy snippets reproducing the report's numeric examples — each one machine-executed and diffed against its printed output before the report ships | |
| ✅ Self-check & glossary | 7 questions (one per step) you should now be able to answer in your own words |
Background knowledge beyond the paper is where LLM reports usually rot. PaperKraken runs a verification protocol on every external claim:
- Cited prior work is summarized only from its actual arXiv/publisher abstract, fetched and read — never from memory.
- Bibliographic details are cross-checked; nothing unfindable is ever cited.
- Math definitions are self-tested with numeric examples before they're included.
- The numeric examples themselves are executed as code (
verify_snippets.py): every "Run it yourself" snippet's stdout is diffed against the number printed in the report, and a mismatch blocks rendering. The worked examples cannot be silently wrong. - Anything that fails verification is dropped, or shipped with an explicit "⚠ unverified — model knowledge" badge. Every verified item carries a source footnote with its check date.
And the PDF itself goes through a mandatory visual QA loop before delivery: pages are rasterized and inspected for broken glyphs, failed math rendering, clipped figure crops, orphaned headings, and near-blank pages. If QA fails, the report is fixed and re-rendered — not shipped.
- Math as SVG via locally bundled MathJax — equations can't fall victim to missing fonts.
- Korean + Latin typography with bundled Pretendard and Satoshi (variable fonts,
word-break: keep-all). - Real pagination via Paged.js: page numbers, running headers showing the current Phase, a cover TOC with live page references, and equation cards that flow across pages row by row instead of leaving half-empty pages.
- PDF outline (sidebar bookmarks) injected after rendering — the full section tree is navigable in any viewer.
- Everything is bundled locally; rendering works offline.
There are many paper-reading skills. They cluster into four families — summarizers, study-material generators, interpretive-article writers, and referee-style reviewers — and each does its one thing well. PaperKraken's claim is the combination: no other skill we found pairs deep comprehension with systematic verification in a print-grade deliverable.
| arXiv summarizers | Study-material generators | Article/slide writers | Referee-style reviewers | PaperKraken | |
|---|---|---|---|---|---|
| Published reading methodology (Nature 7-step) | — | — | — | review frameworks, for referees | ✅ taught to the reader |
| Every numbered equation, term by term | — | rendered, not analyzed | key formulas only | — | ✅ + theorem cards for theory papers |
| Web-verified background with source badges | — | — | — | — | ✅ abstracts actually fetched & diffed |
| Numeric examples machine-executed before shipping | — | code demos, unchecked vs. text | — | — | ✅ mismatch blocks the render |
| Original figures cropped & read element-by-element | — | extracted, not read | partial | — | ✅ with crop QA criteria |
| Critical audit for the reader (independent conclusions, consistency check, COI) | — | — | — | for journal referees | ✅ caught a real 41.0-vs-41.8 discrepancy |
| Print-grade PDF (pagination, running headers, bookmarks, CJK) | markdown | markdown + web viewer | HTML / PPTX | markdown | ✅ + visual QA loop before delivery |
| Korean + English reports | — | docs only | CN/EN | — | ✅ full report in either |
If you want flashcards, slide decks, or runnable re-implementations, those other families are great — and complementary. If you want to actually understand and critically evaluate one paper, with every claim traceable, that's PaperKraken.
Requirements: Claude Code, Node.js ≥ 18, uv, and poppler-utils (pdftoppm, used for QA).
Option A — plugin marketplace (recommended). Inside Claude Code:
/plugin marketplace add devwoo41/PaperKraken
/plugin install paperkraken@paperkraken
Option B — manual, as a personal skill:
git clone https://github.com/devwoo41/PaperKraken.git PaperKraken
ln -s "$(pwd)/PaperKraken" ~/.claude/skills/paperkrakenEither way, one-time renderer deps:
npm install -g playwright-core
npx playwright install chromiumIn any Claude Code session:
> Read this paper with PaperKraken: ~/papers/attention.pdf
> PaperKraken으로 이 논문 읽어줘: https://arxiv.org/abs/1706.03762
You'll be asked one question — report language, Korean or English — and everything else is automatic. The output lands next to the paper:
paperkraken-<paper-slug>/
├── <paper-slug>-report.pdf ← the deliverable
├── report.html ← editable source (re-render any time)
├── figures/ ← verified original crops
└── assets/ ← fonts, MathJax, Paged.js (self-contained)
Works with local PDFs, arXiv links, and (via page-image fallback) even scanned PDFs. ML, experimental-science, theory, and clinical papers each get a type-appropriate methodology audit.
paper.pdf ──▶ full read ──▶ figure extraction ──▶ web verification ──▶ report.html
(type, (caption-detect, (abstracts read, (7-step
equations, crop, visual cross-checked, structure)
number ledger) check, re-crop) badged) │
▼
snippet verification
(numeric examples
executed & diffed)
│
▼
deliverable ◀── visual QA loop ◀── bookmarks ◀── Paged.js + MathJax render
(tofu, math, crops, (Playwright Chromium,
density, orphans) waits for typeset flag)
PaperKraken/
├── SKILL.md # the skill: workflow + QA gates
├── references/
│ ├── methodology.md # Cornwell's 7 steps → report sections; per-type audit checklists
│ ├── equation-analysis.md # equation-card & theorem-card rules
│ ├── verification.md # anti-hallucination protocol for external knowledge
│ └── report-template.md # HTML/CSS skeleton + layout & density rules
├── scripts/
│ ├── extract_figures.py # caption-anchored figure cropping (PyMuPDF)
│ ├── render_pdf.mjs # HTML → PDF, waits for MathJax + Paged.js
│ ├── add_bookmarks.py # injects the PDF outline from report headings
│ ├── verify_snippets.py # executes appendix snippets, diffs vs printed output
│ └── render_pdf.sh # legacy one-shot CLI fallback
├── assets/ # Pretendard · Satoshi · MathJax · Paged.js (all local)
├── docs/ # README previews
└── output_sam/ # sample reports: "Attention Is All You Need" (EN 25 pp · KO 22 pp)
- Methodology: Jacques Cornwell, "Seven steps for critically analysing research papers", Nature Career Column, 2026.
- Rendering: MathJax (Apache-2.0), Paged.js (MIT), Playwright (Apache-2.0), PyMuPDF (AGPL-3.0).
- Type: Pretendard (OFL-1.1), Satoshi (ITF FFL).
If PaperKraken saved you an afternoon, a ⭐ keeps the kraken fed.