Overview
Reading one paper or report thoroughly means jumping between the methods, a results figure, a supplementary table, and an appendix — repeatedly. This skill parses a PDF once into a reusable index of sections, figures, and tables, then answers many questions across the whole document without re-parsing: compare what two sections say, extract every instance of a pattern, and read values off charts/tables.
It is the single-document deep-dive counterpart to the corpus-level tools: to build a
comparison table across many papers use alterlab-pdf-extract; to just convert a PDF
to Markdown use alterlab-markitdown.
When to Use This Skill
Use this skill when the user wants to:
- Interrogate one paper/report end-to-end (methods vs. results vs. appendix).
- Extract every occurrence of a pattern within a document (all p-values, all sample sizes, every mention of a gene).
- Read data off a figure or table inside the PDF.
- Feed precise, located answers into the literature-review or paper-review pipeline.
Does NOT Trigger
| Scenario | Use Instead |
|---|---|
| Build a comparison table across many papers (one row per paper) | alterlab-pdf-extract |
| Convert a PDF/DOCX to clean Markdown | alterlab-markitdown |
| Manage references / DOIs / BibTeX | alterlab-pyzotero |
| Chat with a whole library of sources in a self-hosted NotebookLM-style workspace | alterlab-open-notebook |
| A full multi-reviewer critique of a manuscript | alterlab-paper-reviewer |
Core Capabilities
1. Parse-once, ask-many
Parse the PDF a single time into a structured index — page text with layout, detected section headings, figure/table regions, and (for scans) OCR'd text — then answer subsequent questions against that index instead of re-reading the file each time.
2. Cross-section questions
Answer questions that span the document: "does the discussion's claim match the methods' design?", "which limitations does the appendix add beyond the main text?". Cite the page and section for every answer so it is verifiable.
3. Extract-every-instance
Pull all occurrences of a pattern (every reported p-value, every cohort size, every figure that mentions a variable), returning a located list rather than a single hit — the reliable way to avoid missing one.
4. Read values off figures/tables
Extract tabular data from table regions, and read approximate values from a chart (axis ranges, series values) with an explicit note that chart-read values are approximate and should be confirmed against any underlying data.
5. Pipeline role
Provides the per-document evidence that alterlab-paper-reviewer and literature-review
workflows build on; hand its located extractions to alterlab-pdf-extract when the task grows
from one paper to a whole corpus table.
Resources
references/pdf_explore_usage.md— parser choice and install line, a tested parse-once index + extract-every-instance helper (PyMuPDF), OCR handling, and figure/table reading caveats. Loaded on demand.
Part of the AlterLab Academic Skills suite.