Reading Papers

Reading papers

Get from a fuzzy reference to clean full text with equations intact, cheaply, without wasting context on dead ends. The intelligence is yours; the script is a tool. Don't blindly trust its first answer on hard cases (common author names, paywalled venues). Verify, as flagged below.

The command

~/.claude/skills/reading-papers/scripts/paper.py <cmd> "<query>" [flags]

Flags: --venue "…" · --since/--until YEAR · -n N · --json · --save · --raw · --list-sections · --section "<substr>" · --affiliation "…" · --orcid "…" · --contexts

Keys (optional, all free) live in ~/.claude/secrets/scholar.env; the script auto-loads them.

Run it by path exactly as shown: the uv shebang provisions its dependencies (httpx, lxml, pypdf) in an isolated environment. Invoking it as python3 paper.py bypasses the shebang and fails wherever httpx is not installed (this happened in a subagent environment, 2026-07). If uv itself is missing, fall back to plain curl against the resolver APIs for the one lookup you need.

How resolution works, and why order matters

resolve is cheap and prints whether a free copy exists and where. Always resolve before web-searching or guessing URLs. The routing is cost- and Cloudflare-aware:

• An arXiv id or arxiv link → arXiv metadata (free) + OpenAlex if a DOI exists.
• A DOI → OpenAlex (1 credit) + Unpaywall + NBER direct-PDF + AEA free appendices + a Semantic Scholar OA-PDF fallback (catches NBER copies the others miss).
• An NBER or SSRN URL → mapped to a DOI from the URL itself, skipping the Cloudflare wall.
• A bare title → Crossref first (free); on a confident title match it pulls the record by DOI (1 credit). Only an unconfident title falls through to OpenAlex search (10 credits). So hand it a DOI or arXiv id when you have one; a title costs up to 10× more.

Version-of-record vs preprint: a title search can surface the NBER or SSRN copy. The matcher ranks the journal version above preprint containers, but econ papers genuinely exist as several records with different DOIs and different citation counts, so say which one you mean. Default rule (made explicit after a 2026-07 reading campaign): read the version of record whenever one exists and is reachable; fall back to arXiv/NBER only when it is not, and name the version you read either way, since numbers and author lists can differ across versions.

search is the topic-level entry point (resolve is for a known item). One query per source, merged on DOI → arXiv id → normalized title, re-ranked by reciprocal-rank fusion, one OpenAlex search call per invocation. --json carries abstracts and a sources block saying which sources answered; each record names the source of its citation count. A source that 429s or errors is reported and skipped, never fatal. For a whole literature, use the litreview skill, which wraps this with scoring, Zotero cross-check, and parallel readers.

Read long papers one section at a time

For any paper with an arXiv copy, map first, then pull only what's needed:

paper.py get "10.1257/aer.20181169" --list-sections          # cheap map
paper.py get "10.1257/aer.20181169" --section "decomposition" # that section, raw LaTeX

The single biggest context saver: a 70K-char paper becomes a 10K-char section with every equation, \label, and \ref intact. --section carries the preamble's \newcommand / \newtheorem definitions along, so macros stay resolvable.

Math fidelity ladder

get tries these in order and prints the source: rung it used:

1. arXiv raw LaTeX (--raw, or auto when HTML is absent): the authors' own source. Best for reproducing/checking an equation. Quote the LaTeX verbatim; don't retype it.
2. arXiv HTML (LaTeXML): <math alttext="…"> holds the TeX, macros already expanded, clean prose around it. Best default for discussing a paper.
3. ar5iv: same engine, for papers before arXiv's native HTML (pre-Dec 2023).
4. Born-digital PDF (NBER, JMLR, PMLR, repositories): saved to ~/.claude/cache/papers/; read with ~/.claude/assets/bin/pdfread.py text <pdf> (png for figure pages). Text layer survives; math is fine.
5. OCR: only when get reports (SCANNED). See the OCR rung below.

Rule of thumb: discuss → HTML; reproduce/check an equation → --raw.

The OCR rung: ~/ocr-examples (scanned PDFs only)

get runs a text-layer check and tags a PDF (SCANNED) when the first pages have almost no extractable text. Only then is OCR worth it. If you have such a setup, route it through the ~/ocr-examples repo (an OCR pipeline on an HPC cluster); do not hand-roll OCR.

• Read first: ~/ocr-examples/README.md, docs/ocr-engines.md, and the relevant scripts/*.py --help. Load any HPC plugin skills you use before launching a cluster job.
• Two modes. disk is for public/non-sensitive PDFs (bytes may live on HPC disk). tunnel is for sensitive documents: the Mac reads the PDF and streams pages to a compute-node service over SSH, so bytes never land on HPC storage and the output is written locally. Use tunnel when in doubt about sensitivity.
• Engines (best math first): olmocr2 (olmOCR-2, Apache-2.0, the default), deepseek_ocr, glm_ocr, docling (structure-strong, math-weaker), pypdf (text-native only, no OCR).
• Smallest safe start: one document, one worker, one engine, RTX 8000 not A100. just smoke olmocr2 tunnel proves the path; just documents-process / just engine-olmocr2 run real work. Layout is <documents-root>/<guid>/document.pdf; feed a GUID list via --from-file or direct paths via --pdf-list. Run just sync-hpc after editing that repo.
• Apple-Silicon note: olmocr2/deepseek_ocr/glm_ocr also have a local MLX backend, so a single non-sensitive page can be OCR'd on the Mac without the cluster. A large batch belongs on a GPU node via Slurm, never the login node.

Zotero: the user's citation library

A Zotero MCP server is connected; its tools are underscore-named under mcp__zotero__, e.g. zotero_search_items. This is the user's own reference manager.

• Reads work only while the Zotero desktop app is open with the local API enabled (Settings → Advanced → "Allow other applications on this computer to communicate with Zotero"). If a zotero_* call fails, that's almost always the cause, so ask the user to open Zotero.
• Use it to: search what they already have, read PDF full text and their own annotations/notes, pull BibTeX for citing, and find items by tag. Prefer their copy of a paper over re-fetching; it's faster and it's the version they annotated.
• Writes (adding a found paper to the library) need a web API key in scholar.env and are hybrid-mode. Don't add items unless asked.
• Natural pairing: paper.py finds and reads anything on the open web; Zotero is the private library. When the user says "the paper I saved / my notes on X," reach for Zotero first.

When there is no free copy

Verified July 2026, hard-blocked to plain HTTP: INFORMS (Marketing/Management Science), SSRN, AEA direct PDF, Elsevier/Wiley/OUP/Chicago, OpenReview anonymous api2. Escalate in order:

1. resolve already checked for a green-OA copy (repository, arXiv, RePEc, NBER) and iterates all OA locations, including the ones it doesn't rank first. Trust it; it finds most paywalled AER / Marketing Science papers as a free copy elsewhere.
2. Working-paper version: NBER (nber.org/papers/wXXXXX), CEPR, author's site. For econ this is usually near-identical to the published version.
3. Free AEA appendix/data: for AER/AEJ the article PDF is walled but the online appendix (where the proofs live) and replication package are free; resolve surfaces them.
4. Claude in Chrome, the user's real browser, the only thing that clears Cloudflare. An institutional proxy has the shape https://<your-library>.idm.oclc.org/login?URL=<target>, but login is typically SSO + MFA, so the user must be in the loop. Ask first; one paper at a time. Systematic proxy downloading can get the whole university cut off, so never loop it.
5. Scholar Gateway (semanticSearch): Wiley-leaning licensed corpus, returns passages, not full text; can't fetch by DOI. Good for corroborating a claim, not reading a paper.
6. Say plainly that only the abstract is reachable. Never paraphrase an abstract as if you read the paper.

Cost model (OpenAlex, metered since 2026)

Anonymous = 1,000 credits/day. A free key (openalex.org/settings/api) → 10×; put it in scholar.env as OPENALEX_API_KEY. The script disk-caches every response 30 days, so re-reads are free. Prefer DOIs/arXiv ids over titles.

Author queries: where you must stay in the loop

author resolves a name via OpenAlex (most-works profile) and prints runner-up matches. Common names are genuinely hard; do not trust the top pick blindly:

• OpenAlex fragments one person across profiles: the same author appears as a 300-work record and a 2-work stub, and even an ORCID can resolve to a stub. Always sanity-check that the returned works match the person's field (a marketing scholar shouldn't be returning optics papers). If they don't, you have the wrong fragment.
• --affiliation "Columbia" filters to the current institution, the reliable fix for a common name. If nothing matches, the script lists candidates with their ORCIDs instead of guessing.
• --orcid is the most deterministic single input, but see the fragment caveat above.
• In CS, DBLP is the gold standard for name disambiguation (explicit homonym suffixes) and its author XML also yields the person's Google Scholar / ORCID / ACM ids.
• Diacritics matter: the canonical display_name may be Acemoğlu, not Acemoglu.

cites --contexts shows the actual sentences citing a paper (Semantic Scholar; needs S2_API_KEY). Use it to see what a paper is actually being used for.

Gray areas / known false positives

• Abstract-vs-fulltext: get classifies an HTML page structurally (has a References section? heading count? body word count?), not by raw length, so a short genuine paper with references (an AER P&P note, a Comment) reads as full text and is not flagged. When a page looks abstract-only, get first tries any PDF the page advertises (the citation_pdf_url meta tag, or a linked author manuscript) before falling back, which recovers the full text on many repository landing pages. If it still returns only the abstract, it says so (label ends abstract-only).
• Scanned-PDF flag: heuristic (little text in the first 3 pages). A born-digital paper with a figure-only opening could false-flag; conversely a mostly-image PDF with a text cover could slip through. Glance at the PDF before committing to an OCR job.
• "Open access" that isn't fetchable: for hybrid-OA Oxford/SAGE, OpenAlex reports a PDF URL on academic.oup.com / journals.sagepub.com (legally open, technically Cloudflare-dead). The green-OA (repository/PMC/OSF/arXiv) locations are the ones that resolve; resolve prefers them. Wiley behaves the same: onlinelibrary.wiley.com pdfdirect URLs 403 even for OA articles (confirmed twice, 2026-07), and the landing pages 403 too; the working route is an institutional-repository or author-page copy (MIT DSpace and an author's Harvard page both delivered papers Wiley refused).
• Missing abstracts: Crossref lacks abstracts for JPE (Chicago) and JPSP (APA), though OpenAlex has them. Semantic Scholar elides AEA abstracts by publisher request.

Venue cheat sheet

• CS: almost all on arXiv → --raw gives exact math. Bulk-friendly mirrors are PMLR (per-volume bibliography.bib), ACL Anthology (whole corpus is a git repo), JMLR, NeurIPS proceedings. Reviews/scores need an authenticated OpenReview client.
• Econ: arXiv econ.EM, then NBER (direct PDF). AER/AEJ article PDF is members-only but appendices are free. IDEAS pages give the working-paper ↔ published crosswalk.
• Marketing: INFORMS is walled but green OA, so the readable copy is normally an institutional repository, which resolve finds. 2023+ INFORMS DOIs changed shape; listings contain non-article DOIs (…ack…, …eb…) worth ignoring.
• Psychology: Psych Science is SAGE (PMC holds front matter only). Try PsyArXiv/OSF.
• Annual Reviews: CC-BY articles fetched directly in early July 2026, but as of 2026-07-28 the landing pages sit behind a Cloudflare challenge that blocks curl. Go arXiv-first for AR titles (accepted versions usually exist and worked where the AR page failed); treat any direct-fetch success as date-dependent.

Honesty rules

• State which version you read (published, arXiv preprint, or NBER working paper). The numbers can differ.
• Citation counts are source-specific and disagree (same Econometrica paper: 235 OpenAlex / 177 S2 / 210 Crossref / 157 OpenCitations). Name the source; never mix counts in one table.
• Never invent a DOI, page number, or quotation. If resolve returns nothing, say so.

Quoting the paper

Report what a paper says in indirect speech. Wording kept from the authors goes inside quotation marks and italics, formatted "like this", with the page or section it came from, and that holds in chat, in a subagent's structured summary, in a note field, and in anything that reaches a draft. A summary that reuses the paper's own sentences without those marks is plagiarism once it lands in a document.

Working with many papers

Per the user's standing preference, for several papers at once spawn parallel subagents, one paper each, returning structured summaries. Every API call in paper.py now retries with exponential backoff on 429/503 (arXiv, Semantic Scholar, OpenAlex, Crossref), and the OpenAlex key gives a 10,000-credit daily budget, so a fan-out of a dozen readers is safe. Two things still apply: results are disk-cached (so re-reads across agents are free), and for a very large batch (many dozens) keep arXiv-heavy concurrency modest, since arXiv politeness is ~1 request / 3s.

Setup state

• This setup assumes Zotero MCP (zotero-mcp-launch.sh reads scholar.env), Claude in Chrome, and Scholar Gateway are installed and connected; adjust to your machine. paper.py needs no keys.
• ~/.claude/secrets/scholar.env: put OPENALEX_API_KEY and S2_API_KEY here, plus the optional Zotero web-API creds; REFERENCE.md §7 says how to get all of them.
• Background on the whole landscape (what's blocked, what's open, why a script beat a fleet of MCP servers): REFERENCE.md in this directory.