tyler_med — Medical Literature Wiki Skill

A Claude Code skill that converts a folder of medical / clinical journal PDFs (JVS, JAMA, NEJM, Ann Surg, EJVES, Cochrane, …) into a token-efficient, two-tier Markdown wiki for literature review — with study-design classification, DOI extraction, an evidence-table export, and duplicate detection.

It is a medical-tuned fork of the general-purpose tyler skill from the econtools project (see Acknowledgements).

What it produces

WIKI_DIR/
├── index.md            # papers grouped by study design (evidence hierarchy),
│                       #   with DOI links, sample size (N), and data source
├── index.csv           # evidence-table skeleton (one row per paper)
├── index.json          # same, machine-readable
├── papers/*.md         # cleaned full text + rich YAML frontmatter
├── references/*.md     # trimmed reference lists (kept, not discarded)
└── .wiki_state.json    # incremental cache

Read index.md (or the CSV) to navigate 100+ papers cheaply; open an individual papers/*.md only when you need the full text. The folder is also a ready-to-use Obsidian vault (frontmatter → note properties, [[wikilinks]], nested tags).

Requirements

• Python 3.9+
• pymupdf4llm (bundles PyMuPDF / fitz, used for embedded-metadata and DOI extraction)

pip install pymupdf4llm

Usage

As a Claude Code skill: copy this folder to ~/.claude/skills/tyler_med/ and invoke /tyler_med.

Directly:

python3 convert.py "PDF_DIR" "WIKI_DIR" [OPTIONS]

Flag	Effect
-r, --recursive	Scan PDF_DIR subdirectories
--prefer-pdf-title	Trust the PDF's embedded metadata title over the filename (good for older, badly-named files)
--keep-references	Keep references inline in each paper file
--drop-references	Discard references entirely (default: save to references/)
--force	Re-convert everything, ignoring the incremental cache

Medical-tuning highlights

• Reliable titles from embedded PDF metadata + DOI, not just the filename.
• Mojibake / ligature repair (e.g. n ¼ 4,894 → n = 4,894, dropped fi/fl).
• Study-design classification (RCT, systematic review, meta-analysis, cohort, case-control, guideline/consensus, scoping/narrative review, QI, protocol, editorial). A named registry in the title (NSQIP, VQI, Medicare, …) implies a cohort; an explicit EDITORIAL tag in a title is honoured.
• Structured metadata: DOI, sample size (n=), data source, trial/PROSPERO registration IDs, and clean controlled tags.
• Duplicate detection (shared DOI or year+title) with ⚠️DUP-n flags.

Note: study design, sample size, and data source are heuristic (best-effort from the abstract). Verify anything load-bearing against the full text.

Using it as an Obsidian vault

Open WIKI_DIR in Obsidian (or drop it into an existing vault). Each papers/*.md carries YAML frontmatter that becomes note properties, and the index links every paper with [[wikilinks]].

Tags are deliberately minimal and controlled — only three nested namespaces, so the tag pane and graph stay clean and queryable:

• design/… — study design (e.g. design/cohort-study, design/randomized-controlled-trial)
• source/… — data source (e.g. source/nsqip, source/vqi, source/medicare/cms)
• year/… — publication year (e.g. year/2023)

Per-keyword hashtags are off by default (they flood the graph); the human-readable keywords: property is kept regardless. Pass --keyword-tags if you want them.

Dataview turns the frontmatter into a live evidence table. Examples:

```dataview
TABLE study_type AS Design, year AS Year, sample_size AS N, journal
FROM "papers"
WHERE contains(data_source, "VQI")
SORT year DESC
```

```dataview
TABLE WITHOUT ID link(file.link, title) AS Paper, doi
FROM #design/randomized-controlled-trial
SORT year DESC
```

For a by-topic view (what to cite for a given point), ask Claude to build index_by_theme.md (see Step 4 in SKILL.md) — a curated thematic grouping with a design badge and one-line contribution note per paper.

Acknowledgements

This project is a derivative work of the tyler skill in the econtools project by @johanfourieza, used and continued under the MIT License. The original tyler was built for economics / social-science papers; tyler_med re-tools it for clinical literature (study-design classification, DOI/registry metadata, evidence-table export, mojibake repair, duplicate detection). With thanks to the original author.

License

Released under the MIT License — see LICENSE. Copyright is held by johanfourieza for the upstream tyler skill and by Robert A. Larson, MD for the medical adaptation.