Paper Reading

Paper Reading — Translation + Critical Review

This skill turns a scientific paper PDF into two Korean markdown files following a fixed, repeatable workflow. The goal is a faithful full translation plus a genuinely critical review — not a shallow summary. LLMs reliably capture contributions and methods but tend to under-produce substantive criticism, so the review steps below are scaffolded deliberately to force real weakness-finding.

Output: two files in a dedicated folder

For every paper, create one folder and two markdown files. Never leave the source PDF loose in the parent directory.

Naming

• ID: {firstAuthorLastname}{year}{keyword} (lowercase), e.g. yang2024boron
• Folder: {ID}_{journalAbbrev}, journal abbreviation in CamelCase, hyphens allowed, no spaces,

e.g. yang2024boron_ACSAppliedMat-Inter

17_paper_reading/
└── yang2024boron_ACSAppliedMat-Inter/
    ├── yang2024boron.pdf              # source PDF, moved + renamed here (mv, not cp)
    ├── yang2024boron_translation.md   # Step 1
    └── yang2024boron_review.md        # Steps 2–5

Move and rename the PDF into the folder so no original copy remains in the parent.

First-run setup: the default working root is stored in config.local.json next to this SKILL.md ({"workspace": "/path/to/paper_reading_root"}). If that file does not exist yet, ask the user once: "논문 정독 폴더들을 모아둘 기본 경로를 알려주세요 (예: ~/papers). 앞으로 PDF를 주시면 그 아래에 폴더를 만들어 정리합니다." Save the answer and don't ask again. If the user is clearly working in a specific directory for this paper, that takes precedence.

Before you start: read the PDF properly

Use the pdf skill (or the user's available PDF tooling) to extract text — do not guess at content. Read the entire paper including Methods/SI prose, not just the abstract and figure captions.

PDF text extraction routinely mangles equations and multi-column layouts. When you hit a broken equation or garbled table, flag it rather than inventing content, and note it in the Step-5 self-review. Figure/Table captions must still be translated in full; the figures themselves are images you cannot read from extracted text, so reason from the caption + surrounding prose.

Step 1 — Full line-by-line translation (*_translation.md)

Translate every sentence of the paper, English original and Korean rendering in parallel. Preserve the paper's exact section/subsection structure. Include every figure and table caption.

Open the file with a short metadata block and a key-term glossary, then proceed section by section. The glossary locks one Korean rendering per recurring technical term so terminology stays consistent across a long paper (and doubles as a study aid). Build it as you translate; keep established acronyms in English (MLIP, DFT, NEB, …).

# {Paper title}

**Authors**: ...
**Journal**: {name} {vol}, {pages} ({year})
**DOI**: https://doi.org/...

---

## 용어집 (Key Terms)

| English | 한국어 | 비고 |
|---------|--------|------|
| adsorption energy | 흡착 에너지 | |
| selective task regularization | 선택적 태스크 정규화 | 본 논문 핵심 개념 |

---

## {Section Title} / {섹션 제목}

> Original English sentence.

한국어 번역 문장.

> Next English sentence.

다음 한국어 번역 문장.

Translation quality matters more than speed:

• Keep technical terms accurate and consistent across the whole paper (pick one Korean rendering

per term and stick to it; keep established English acronyms as-is, e.g. MLIP, DFT, NEB).

• Translate faithfully, not loosely — this is a study aid, the user will compare against the original.
• For a very long paper (30+ pages), translating dense Methods detail (full loss-function

derivations, exhaustive DFT settings) can be compressed to its core, but say so explicitly in the self-review. Main-text prose is always translated in full.

Step 2 — TL;DR + claimed strengths (*_review.md)

Open *_review.md with a TL;DR — a 3–5 line Korean digest a reader can skim instead of the whole review. Answer: what problem, what method, the headline result (with the single most important number), and the one caveat worth knowing.

## TL;DR

- **문제**: ...
- **방법**: ...
- **핵심 결과**: ... (가장 중요한 수치 1개)
- **주의할 점**: ...

Section 1 — Claimed strengths

List the contributions and strengths the authors themselves claim. For each item, cite the backing experiment / number / figure or table. This is the part LLMs do easily — keep it tight and evidence-anchored, no padding. Bullet list.

Step 3 — Core logic / key arguments (*_review.md, section 2)

Reconstruct the paper's reasoning chain: hypothesis → method → experimental design → conclusion. Write it as "why this method → what it showed → what they concluded." Anchor each logical step to its Figure/Table number and key numbers. A short text-based flow diagram is encouraged when the argument has a clear pipeline.

Step 4 — Limitations & critical notes (*_review.md, section 3)

This is the highest-value section and the one models tend to shortchange. Push hard here. Separate into two categories:

• 저자 인정 한계 (author-acknowledged): limitations explicitly stated in the paper, quoted or

paraphrased with location.

• 추가 비판 (additional critique): weaknesses you find by reading critically. Probe each of:

experimental design and fair-comparison concerns; missing baselines or ablations; hyperparameter arbitrariness and reproducibility; generalization beyond the tested architecture/dataset; unreported training/compute cost; data-quality issues; scope gaps (what the stated motivation promised vs. what was actually delivered); over-claims where reported gains ≠ physical/real-world significance.

Be specific and grounded — "no compute cost reported despite a 250M-structure model" beats "could be more efficient." Aim for a substantial list; a paper with only one or two critiques usually means you did not look hard enough.

A short overall assessment (contribution, predicted impact, one-line verdict) is a good closer but optional.

Step 5 — Self-review (*_review.md, final section)

After Steps 1–4, re-read the paper from start to finish against what you wrote and verify:

• No missing or mistranslated passages.
• No claims in the strengths/logic/limitations that contradict the source.
• Every Figure/Table number cited is correct.
• Every performance number (accuracy, error, dataset size, hyperparameter value) matches the original.

Fix errors immediately, then record what you checked, any potential errors found, and what was intentionally compressed. Save both files as final.

Worked examples

If the user's workspace already contains finished paper-reading folders from previous runs, read one before starting when you need a concrete target for depth and format.

Optional enhancements

TL;DR and the key-term glossary are now part of the default workflow (Steps 1–2). The following remain off by default to keep output focused; offer or apply them when the user wants more. See references/enhancements.md for details on each:

• Prior-work comparison table
• Reproducibility checklist
• Extracting figures as image files for visual reference

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: JinukMoon
• Source: JinukMoon/agent-skills
• License: MIT

Install and usage instructions live in the source repository linked above.