Allpapers

allpapers

An agent skill for finding and retrieving the full text of scientific papers. It installs into Claude Code, the Codex CLI and Antigravity (`agy`) — see [Install](#install).

It merges the older `paperclip` and `scihub-cli` skills, adds CORE and Unpaywall, adds arXiv LaTeX source retrieval, and arranges every source into one priority ladder. The organizing rule is that **a parseable text format is worth more than a convenient one**: LaTeX and JATS XML carry equations, section structure and reference lists that PDF text extraction destroys.

The three rules

Everything below follows from these. They are ordered: when two pull against each other, the earlier one wins.

1. Prefer parseable source. PDF text extraction silently mangles equations, loses section boundaries, splits ligatures and reorders multi-column text. LaTeX and JATS do not. Reaching for a PDF when LaTeX exists means quoting from a worse copy for no reason.
2. Follow the ladder. Cheap, open, already-extracted sources first; bootleg copies only when nothing else has it. Do not skip rungs to reach the bottom faster — the lower rungs are slower, less reliable, and in one case not open at all.
3. Nothing is cited until it is recorded. Any paper supporting a claim goes into verification/bib.md with its composite BibTeX entry, every source URL used, its abstract, a justification, and verbatim quotes with locators. Papers examined and rejected get an entry too: knowing a dead end was already explored is worth as much as an inclusion.

Rule 1 is enforced in code. `allpapers-locate` sorts every location it finds by `FORMAT_RANK`, best first:

extracted text  ›  LaTeX  ›  XML  ›  HTML  ›  PDF  ›  unknown

Within the XML tier, prefer authored JATS over a GROBID machine parse of a PDF — the two share a rank in the code but not in quality. A scanned PDF with no text layer is the worst case of all and needs reading by eye.

Install

The skill is