shzhao27208/aut_sci_write/sci-extract

Sci-Extract — Scientific Extraction

Professional extraction of core insights and figures from scientific PDF papers.

Note: This skill includes contributions from two authors. See Copyright & License section for details.

Features

• Core Insights: Automatically identify research problem, methodology, key results, innovations, applications, and limitations.
• Review Literature Extraction: For reviews, surveys, systematic reviews, scoping reviews, and meta-analyses, extract the field scope, taxonomy, evidence base, consensus, disagreements, evidence quality, research gaps, and future directions.
• Paper Harvesting: Given a DOI, PMID, PMCID, arXiv ID, title, or search query, pull complete metadata from every configured database, the native publisher XML, the full body text, and the figures, then write an agent-readable raw.md instead of leaving you to read a PDF.
• Figure Detection: Fetch publication-resolution figures from the publisher, or crop figure regions from PDF pages when the publisher ships none.
• Metadata Extraction: Parse title, authors, DOI, journal, and year.

Steps：

Step 1: Acquire the paper

Always read the paper fresh. Never rely on memory of the paper, even if the title looks familiar.

If the paper is long, first classify the paper type, then prioritize the sections relevant to that type. For original research, prioritize abstract, introduction, method/theory, experiments, conclusion. For review literature, prioritize abstract, introduction, search/selection methods, taxonomy/classification sections, major thematic sections, summary tables, limitations, and future perspectives.

Step 2: Classify the paper type

Before choosing the extraction template, classify the paper as one of:

• Original research article
• Narrative review / survey
• Systematic review
• Meta-analysis
• Scoping review
• Perspective / tutorial / methods overview

Use the title, abstract, introduction, and section headings. Signals for review literature include "review", "survey", "systematic review", "meta-analysis", "scoping review", "bibliometric", "taxonomy", "current status", "recent progress", "challenges", "future perspectives", and broad comparison tables.

If the paper is a review, survey, systematic review, meta-analysis, scoping review, perspective, or tutorial overview, do not force it into the original-research Heilmeier template. Use the Review Literature Extraction Mode below. Original research asks "what did this paper do and prove"; review literature asks "how does this paper map the field and judge the evidence".

Step 3A: Original Research Mode - Answer the modified Heilmeier questions

Answer each of the seven questions below as a labeled subsection, in order. For each question, the rules differ on (a) whether your own evaluation is allowed and (b) whether external citations are allowed. Read the rules carefully before writing each subsection.

Question 1. What are you trying to do?

Open with a one-sentence statement of the paper's contribution written for a smart non-specialist, with absolutely no jargon. Ban acronyms and any technical term a first-year undergrad would not know. If a term of art is unavoidable, define it parenthetically in plain words. Then add one or two sentences expanding the objective in slightly more technical language.

Opinions allowed: no. Stay faithful to the paper. External citations allowed: no.

Question 2. What is the problem, how is it done today, and what are the limits of current practice?

Describe the real-world or scientific problem the paper addresses, then give a brief overview of how the field handles it at the time of the paper, and what the limitations are. This is meant to be a self-contained landscape paragraph, not a literature review. Cover the main competing approaches in plain prose.

Opinions allowed: a small amount, only if it sharpens the framing of the limits. External citations allowed: no. Do not search for or cite outside sources here. Just give an overview from the paper and your general knowledge of the field.

Question 3. What is new in the approach, including core idea, math, and method, and why does the paper claim it will succeed?

This is the technical heart of the response and absorbs what would otherwise be a "method" summary. Cover, in this order:

1. The central technical move that distinguishes the paper from prior work.
2. The key mathematical objects and formulation. Include the main equation or two, define every symbol you introduce, use display math with \left and \right for brackets, keep inline math on one line, and prefer standard LaTeX notation.
3. How the proposed method actually solves the problem mechanically.
4. The paper's own claim about why the approach will succeed.

Opinions allowed: NO. This subsection is strictly about what the paper says and proposes. Save your evaluation for questions 4, 5, and 6. External citations allowed: no.

Question 4. Who cares? If successful, what difference does it make?

Discuss the impact: which communities benefit, what becomes possible, and whether this paper has actually shifted the field since publication.

Opinions allowed: yes. This is one of the questions where your judgment matters most. External citations allowed: yes, and encouraged when assessing post-publication impact (adoption by other groups, follow-up papers, deployment). Every external citation must come from a web_search or web_fetch you actually ran in this turn.

Question 5. What are the risks?

Cover both the risks the paper itself acknowledges and the ones you see independently. Be concrete: contamination, reward hacking, failure modes, narrow benchmarks, scaling, reproducibility.

Opinions allowed: yes. External citations allowed: yes, when an outside source materially supports a risk claim.

Question 6. How much will it cost?

Interpret as compute, data, engineering effort, or deployment cost, depending on the paper. State which interpretation you are using. Pull whatever numbers the paper provides (token counts, batch sizes, GPU hours, data volumes) and translate into a rough sense of "what would it take to reproduce this".

Opinions allowed: yes, especially for the "what would it take to reproduce" framing. External citations allowed: yes. Be careful not to conflate this paper's costs with related work by the same authors. If you cite a cost figure, state exactly which paper or model that figure refers to.

Question 7. What are the experiments and results?

Cover the experimental setup (benchmarks, datasets, baselines, metrics, ablations) and the headline results. This subsection answers "what are the criteria for success and did the paper meet them". Note any conspicuous gap between claims and evidence.

Opinions allowed: small amount, only for noting gaps between claims and evidence. External citations allowed: no.

Step 3B: Review Literature Extraction Mode

Use this mode for review papers, survey papers, systematic reviews, meta-analyses, scoping reviews, perspective reviews, and tutorial overviews. Do not ask for a single proposed method, single experiment, or reproduction cost unless the review itself is about benchmarking or a method protocol. The goal is to reconstruct the paper's map of the field and assess how reliable that map is.

Return the following labeled subsections, in order:

1. Review scope

State what field, problem, population, method family, material class, data type, or application area the paper reviews. Include explicit inclusion and exclusion boundaries when the paper gives them. If the scope is vague, say so.

2. Why this review exists now

Explain the motivation: field fragmentation, rapid growth, conflicting evidence, unclear taxonomy, translation gap, reproducibility problem, new technology, or practical need. Keep this faithful to the paper before adding your own judgment.

3. Taxonomy or organizing framework

Extract the categories the authors use to organize the field. Preserve the author's hierarchy and terminology. If there are multiple taxonomies, separate them. For each category, give a one-sentence meaning and the main representative approaches, study types, models, materials, datasets, diseases, interventions, or applications.

4. Literature selection and evidence base

For systematic reviews, scoping reviews, and meta-analyses, extract databases searched, search period, search terms if available, inclusion criteria, exclusion criteria, screening process, final included study count, and any quality-assessment tool. For narrative reviews and surveys, state whether the search and selection method is unspecified or informal.

5. Field landscape and main themes

Summarize the major research directions covered by the review. Distinguish mature areas from emerging areas. Mention important datasets, benchmark tasks, experimental platforms, clinical cohorts, model families, materials, or instruments when they are central to the review.

6. Consensus findings

Extract what the reviewed literature broadly agrees on. Separate strong consensus from tentative patterns. If the review does not clearly identify consensus, say so instead of inventing one.

7. Disagreements, controversies, and heterogeneity

Extract where studies conflict, where mechanisms or interpretations differ, and what explanations the review gives for inconsistent findings. For meta-analyses, include heterogeneity statistics and subgroup/sensitivity findings if reported.

8. Evidence quality and bias

Assess the reliability of the reviewed evidence using what the paper reports: study design, sample size, data quality, benchmark leakage, confounding, publication bias, reproducibility, missing controls, annotation quality, evaluation metrics, or risk-of-bias tools. Mark your own assessment with a first-person phrase such as "My analysis is that," so paper content and your critique remain separate.

9. Gaps and future directions

Extract the open questions, missing datasets, missing experiments, technical barriers, translation barriers, standardization needs, clinical/industrial/policy needs, and concrete future directions identified by the authors. Add your own prioritized gap assessment only with a first-person marker.

10. Important figures and tables

Identify taxonomy figures, workflow diagrams, evidence maps, PRISMA flow diagrams, summary tables, comparison tables, benchmark tables, and meta-analysis forest/funnel plots. For each important figure or table, explain what role it plays in understanding the review. If figure extraction is requested, crop or save the relevant figure/table regions when tooling is available.

11. Bottom-line takeaways