Coarse Extract

/coarse-extract — Extract and Parse Structure

Usage: /coarse-extract papers/paper.pdf

Argument ($ARGUMENTS) is the path to the paper. All paths are relative to the workspace root d:/Dropbox/Research/Coarse Reviewer/.

Path setup

Derive a slug from the filename: lowercase, spaces→hyphens, no extension. e.g. papers/my-paper.pdf → slug = my-paper

• PAPER_PATH = $ARGUMENTS
• EXTRACTED = .coarse_cache/_extracted.md
• SECTIONS_JSON = .coarse_cache/_sections.json

Create .coarse_cache/ if it doesn't exist.

---

Step 1 — Extract (Python)

Run:

python scripts/coarse_extract.py PAPER_PATH EXTRACTED

This writes the full paper as markdown to EXTRACTED. Uses Mistral OCR via OpenRouter if OPENROUTER_API_KEY is set; otherwise falls back to pymupdf.

If extraction fails, try the pymupdf fallback directly:

python -c "
import fitz, sys
doc = fitz.open(sys.argv[1])
text = '\n\n'.join(page.get_text() for page in doc)
open(sys.argv[2], 'w', encoding='utf-8').write(text)
" PAPER_PATH EXTRACTED

Then read EXTRACTED into your context with the Read tool.

Garble check: After reading EXTRACTED, scan the first 3000 characters for OCR artifacts — spaced-out characters (e.g. T h e o r e m), garbled symbols (ˆ b, ∈ˆ, HTML entities like <), or lines that are clearly noise. If you find more than a few isolated artifacts, warn the user: the PDF may have been scanned or image-only. The downstream review steps include an OCR-artifact notice to the reviewer, but severe garbling will hurt review quality. Advise the user to check whether a text-native PDF is available.

---

Step 2 — Parse Structure (Python)

Run:

python scripts/coarse_structure.py EXTRACTED SECTIONS_JSON

This writes a JSON array of section objects to SECTIONS_JSON. Each section has:

• number, title, text, section_type (abstract/introduction/related_work/

methodology/results/discussion/conclusion/appendix/references/other)

• math_content (bool — set to false by this script; will be set in classify step)
• claims (list of theorem/lemma/proposition statements found by regex)
• definitions (list of definition statements found by regex)

Read SECTIONS_JSON and confirm the section list looks correct.

---

Done. Next: /coarse-classify

Source & license

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

• Author: Felipe-SO
• Source: Felipe-SO/coarse-ink-claude-code
• License: MIT

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