drpwchen/paper-review-and-digest

Two Claude Code skills for reading a single
academic paper properly — one appraises whether it can be trusted, the other absorbs what it says.
Built by a practising physician for real journal-club and clinical use, then generalized for sharing.
🔭 Discovery paper-radar ｜ ⬇️ Download paper-fetch ｜ 📖 Reading paper-review-and-digest — a literature radar finds and ranks papers, a route-ladder fetcher gets their full text, and these skills appraise and digest them; together the three repos form the full paper pipeline.
> English · 繁體中文
📝 How it works and why, in plain language: the blog post.
## What this is
Reading a paper well is really two different jobs, and good readers keep them separate:
| Skill | Answers | What it does |
| /paper-review | Can I trust this? Is it done well? | Journal-club-grade appraisal: journal credibility, author track record, reference verification with a live CrossRef existence gate, design-routed risk-of-bias, spin / argument-structure audit, deterministic GRADE, citation-impact analysis. |
| /paper-digest | What does it say? How do I absorb it fastest? | A teaching-style content digest in three progressive-disclosure layers, structure routed by paper type, ending in active-recall self-test cards. Hard-stops without full text (no abstract-only fake digests). |
They are independent and can both run on one paper.
> Part of a pipeline. These skills are the reading end. The discovery end — a personal
> literature radar that pulls dozens of journal feeds + PubMed searches, ranks them by your interest
> model, and hands the picks off for appraisal/digest — is
> paper-radar. paper-radar's /paper-sync dispatches
> a picked paper straight into /paper-review (🔬 quality) or /paper-digest (📚 content). Use them
> together, or either alone.
## Why it isn't just a clever prompt
Most "AI paper review" tools make two claims loosely — that they grade the evidence and that they
catch spin. Here those claims are made true by real, deterministic code. The split is the whole
point: the language model does semantic judgement; a program does the logic and the arithmetic.
### grade_judge.py — GRADE as arithmetic, not vibes
The model rates the five GRADE domains (risk of bias, inconsistency, indirectness, imprecision,
publication bias) as not serious / serious / very serious. A pure-stdlib script then recomputes
the final certainty from the starting level and the summed downgrades/upgrades, clamped to
[Very low … High]. The model's self-reported grade is advisory only; the script's number is
authoritative, and it warns when the two disagree. It also enforces GRADE rules the model tends to
forget — e.g. observational evidence can't be upgraded once any domain is downgraded.
### argdown_lint.py — inference gaps as a real gate
The model tags each supporting finding by premise type (direct_rct, association,
surrogate_outcome, single_study, subgroup, secondary_outcome, mechanistic,
expert_opinion). A stdlib script then deterministically flags the illicit jumps and exits
nonzero:
surrogate outcome → hard-endpoint / causal benefit claim
association → causal claim
single study → "consistently shown"
subgroup / secondary outcome only → benefit claim (classic spin)