dangeles Papers Library
Announce: "I'm using the papers-library skill."
The library lives at ~/repos/papers. Read ~/repos/papers/CLAUDE.md before acting —
it holds the full protocol. ~/repos/papers/RECIPE.md holds the summary format. This skill
exists so those are reachable from any repo; it does not duplicate them.
There is no bare python on this machine. Run the tools through the repo venv, which works
from any directory:
~/repos/papers/.venv/bin/python ~/repos/papers/tools/validate.py
Which flow are you in?
The user handed you a paper ("add this", "summarize this for the library", a bare DOI or
PDF path) → ingest. Resolve the DOI, fetch metadata from OpenAlex rather than transcribing
it, work down the fetch ladder in CLAUDE.md until you have the best artifact available,
write the sections per the recipe, run tools/hmp.py, then tools/validate.py. If what you
read was not a PDF held in pdfs/, say which artifact it was under ## Full Text Source.
The user wants an entry checked ("does this still hold", "check these against their PDFs") → verify. Re-run the fetch ladder, read the new artifact, and correct the entry against it — figure panels first, since a rendering hides the p-values printed on them. Every factual correction this library has made came out of that step.
The user is writing and wants a citation ("cite Igarashi for the OSM point") → this is
the flow that makes the library worth having. Insert the citation into the manuscript in its
native format and append the Cited For entry to the library file in the same step, with
cited_by_count and cited_by_sources updated to match — the validator errors when they
disagree. Don't defer it. If the paper isn't in the library, create a stub and say so.
The user points at a manuscript ("scan my grant") → tools/scan.py <path>. Idempotent;
safe to re-run.
The user wants an entry changed ("apply the requests on X", "fix the stats block") → edit
in place, bump updated:, leave ## Cited For untouched, and re-run hmp.py if the stats
block moved.
The rules that matter most
Never hand-write hmp or flags — run tools/hmp.py <file> on the named file. A guessed
harmonic mean p-value looks like knowledge and isn't, and validate.py will catch it anyway.
Never delete anything under ## Cited For. It is append-and-update only.
Never crawl reference lists. The user curates this library. Reading a paper may surface at
most two papers its own authors flag as essential; record them in suggested: and create
nothing.
Never claim to have read a full text you couldn't get — that's what status: abstract-only
is for.
Finish by running .venv/bin/python tools/validate.py in the repo. Zero errors, or the work
isn't done.
Finding things
grep -rl "oncostatin" ~/repos/papers/papers/
grep -rn "equivalence_from_nonsignificance" ~/repos/papers/papers/
Adjacent skills
paper-reader— understanding a paper without recording it in the library.researcher— extracting parameters with citations for a model or calculation.lit-pm— a comprehensive review across many papers.