v3.0.0 is out

Turn one dense paper into an Obsidian note worth keeping.

DeepPaperNote is an agent skill that runs in Claude Code, Codex, and any other agent that supports skills. Hand it a paper, and it reads the full text, writes an illustrated deep-reading note, saves it into your Obsidian vault, and links it to the other papers you have read. Every paper you read grows your academic wiki.

Why DeepPaperNote

From one sentence to a note you can trust and actually use.

Effortless

One sentence, one note

No prompts to learn, no templates to set up. Give it a title, DOI, arXiv ID, or PDF, and it finds the paper, reads it in full, extracts the figures, and saves the note into your vault.

Expert

Reads like a domain expert

From CS and AI to medicine, psychology, and the humanities, each of 10 paper types gets its own way of reading, drawn from more than 50 reviewer guidelines and reporting standards and tested on 34 real papers.

Trustworthy

Accurate and complete

Every conclusion traces back to the paper, numbers come straight from its text and tables, and every main experiment is covered. Each note passes an independent review and a strict quality check before saving, and when the full text cannot be read, it never makes one up.

Visual

Figures, tables, and formulas in the text

Figures and tables are cropped from the PDF and placed where the note discusses them, and formulas render in Obsidian. No more flipping back to the PDF.

Your vault

A vault that knows you better the more you use it

Each note links to related papers in your vault and goes into the domain folders you already use. Record your research directions, and each note's Takeaways tell you what the paper means for your own work.

The output is the product.

Every claim in the note traces back to a page, a figure, or a table in the paper.

Figures sit right next to the explanation

Paper Figure 1: the Transformer model architecture
Paper · p. 3 · Figure 1
Note Method › Method Workflow

Original paper item: Figure 1. The complete encoder-decoder architecture, including shifted target inputs, positional additions, residual paths, and the output probability head. Source p. 3.

Training and generation use the same conditional prediction structure with different target inputs. During training, shifted reference tokens provide the preceding targets, a practice commonly called teacher forcing. All known target positions can then be processed together under the causal mask. During generation, the next target is unknown until the decoder predicts it, so generation still proceeds one token at a time. [Sections 3 and 6.1]

Numbers from the paper, contradictions included

Paper Table 2: BLEU scores and training cost
Paper · p. 8 · Table 2
Note Experimental Evidence › Translation quality
Model from Table 2English-German BLEUEnglish-French BLEU
MoE26.0340.56
ConvS2S ensemble26.3641.29
Transformer base27.338.1
Transformer big28.441.8

The French result requires an explicit source caveat. The abstract and Table 2 report 41.8, but the prose of Section 6.1 reports 41.0. Both exceed the listed best earlier single-model score of 40.56, yet only 41.8 exceeds the listed 41.29 ensemble.

Not just the numbers: what they actually show

Paper Table 3: ablations of the Transformer architecture
Paper · p. 9 · Table 3
Note Experimental Evidence › Component ablations
FamilyChange from base, unless statedDev perplexityDev BLEU
Reference6 layers, width 512, 8 heads, key and value width 644.9225.8
A1 head, key and value width 5125.2924.9
DDropout 05.7724.6
DLabel smoothing 04.6725.3

Removing dropout worsens both development perplexity and BLEU. Removing label smoothing improves perplexity from 4.92 to 4.67 but worsens BLEU from 25.8 to 25.3, showing why probability fit cannot replace the task metric. [Table 3 B-D]

From Paper to Wiki

DeepPaperNote reads each paper deeply and saves it well. Three optional companion skills look after your vault as a whole.

deeppapernote Core

Deep-reads one paper and saves an illustrated note into your vault, linked to the other papers you have read.

<your papers folder>/
└── NLP/
    └── Attention_Is_All_You_Need/
        ├── Attention_Is_All_You_Need.en.md
        └── images/
            ├── figure_1.png
            └── table_2.png

/deeppapernote “Generate a deep-reading note for this paper: <paper>”

research-memory

Records the research directions you are working on and follows them as you read. With it, each note's Takeaways are written against your directions, so the next note knows you better.

/research-memory “Build my research memory”

organize-wiki

Sorts out your domain categories, moves papers filed under the wrong domain, and fixes their tags. The categories it settles guide where new papers are filed.

/organize-wiki “Organize my paper library”

paper-glossary

Turns a paper's key concepts into reusable term notes linked back into the text. The same term in different papers points to the same note.

/paper-glossary “Build glossary notes for this paper”

  1. Read a few papers
  2. Once notes pile up, build a research memory
  3. When categories drift, organize the vault

Install the skill. Hand it one paper.

A title, DOI, arXiv ID, or local PDF is enough. The first time, it asks only two things: which language to write in, and where to save.

Already installed? See how to update →

# 1. Install

claude plugin marketplace add 917Dhj/DeepPaperNote claude plugin install deeppapernote@deeppapernote

# 2. Install the PDF dependency

python3 -m pip install PyMuPDF

# 3. Read a paper

/deeppapernote https://arxiv.org/abs/1706.03762

Thanks for reading, using, and supporting DeepPaperNote. May your paper-reading days be a little clearer, calmer, and more rewarding.