An AI skill that teaches you how to read and evaluate scientific papers. Built for OpenClaw and Claude Code.
Drop in a scientific paper (PubMed link, DOI, PDF, or pasted text) and the skill walks you through it section by section using a proven methodology — then evaluates the study's quality and delivers a plain English verdict.
Two modes:
- Guided (default) — Walks you through each section with questions. Asks what YOU think before giving its assessment. Builds your ability to read papers independently.
- Fast (
--fast) — Skips the walkthrough, delivers the verdict directly.
Based on Dr. Jennifer Raff's "How to Read and Understand a Scientific Article" — a structured three-pass approach:
- Orientation — Abstract, figures, hypothesis (2-3 min)
- Deep Read — Introduction, methods, results, discussion (15-30 min)
- Critical Evaluation — Form your own interpretation before reading others' (5-10 min)
Every paper is scored on six criteria:
| Criteria | What It Checks |
|---|---|
| Sample Size | Adequate, marginal, or inadequate for the claims being made |
| Funding & COI | Who paid, any conflicts of interest |
| Methodology | Where on the evidence hierarchy (meta-analysis > RCT > cohort > case report) |
| Effect Size | Practically meaningful or just statistically significant |
| Applicability | Does the study population match your context |
| Confidence | 1-10 score based on cumulative assessment |
Output: A plain English verdict — Solid / Weak / Inconclusive / Do Not Apply — with a confidence score and every flag explained.
- Health / supplements / protocols — Confidence threshold 8. Peer-reviewed required. No study backing a claim = hard flag.
- Training / exercise science — Confidence threshold 8. Inherent blinding limitations noted.
- General science — Confidence threshold 7.
Copy the study-reader folder to your skills directory:
cp -r study-reader ~/.codex/skills/OpenClaw auto-discovers skills from ~/.codex/skills/. Restart your gateway service to pick it up.
Copy to your Claude Code skills directory:
cp -r study-reader ~/.claude/skills/For Claude Code, add argument-hint and allowed-tools to the SKILL.md frontmatter:
argument-hint: <paper-pdf-or-url> [--fast]
allowed-tools: [Read, Glob, Grep, Bash, WebFetch, WebSearch, Agent]Input: PubMed link to Douris et al. (2015) — "Martial Art Training and Cognitive Performance in Middle-Aged Adults"
Output:
Verdict: Inconclusive
Confidence: 5/10
Study type: Repeated measures experimental (single session)
Sample: N=10, all experienced black belts in Soo Bahk Do, mean age 53.5
Funding: Not disclosed
What's strong:
- Well-designed within scope (randomized, counterbalanced)
- Validated cognitive test (Stroop)
- Large effect sizes (d=1.1-1.8)
What's weak:
- N=10 is very small
- All participants already trained — nothing about beginners
- Single session, not longitudinal
- Intensity not matched between conditions
Bottom line:
Interesting signal, not proof. Raises a good question but doesn't answer it.
This skill was built on the work of:
-
Dr. Jennifer Raff — "How to Read and Understand a Scientific Article". The three-pass reading methodology and Socratic question framework are adapted from her guide. Dr. Raff is an Associate Professor of Anthropology at the University of Kansas and author of Origin: A Genetic History of the Americas.
-
Ramonov Quality Rules — The confidence scoring and flag system is inspired by the Sabrina Ramonov principles for AI output quality: no guessing (flag uncertainty explicitly), confidence scores on every finding, and verified sources listed separately.
-
Evidence Hierarchy — The methodology ranking (systematic review > RCT > cohort > case-control > cross-sectional > case report) follows standard evidence-based medicine frameworks as taught in medical and nursing education.
Rob / OldManStillCan — a 55-year-old t-shirt maker who built this to help his wife read research papers in nursing school. Not an expert. Just inspired by what these systems make possible.
Built with Claude Code (Hephaestus).
MIT