Connected Papers Skill — Development skill for Claude Code

Ask *"what is this paper built on?"* or *"who built on it?"* and get an answer grounded in a real citation graph instead of model recall — with the graph's own traps handled for you.

This wraps the [Connected Papers API](https://github.com/ConnectedPapers/connectedpapers-py): resolve whatever paper reference you have into an ID, fetch the neighborhood, and turn it into prose that answers the question you actually asked.

It ships as two independently useful pieces:

• scripts/cp_map.py — a plain Python CLI. Standard library only for resolve, s2map, and report; no agent, no framework, no vendor SDK.
• SKILL.md — the hard-won judgment about reading the graph correctly, written as instructions any capable model can follow.

Why this exists

Connected Papers returns a rich graph, and the naive readings of it are wrong in ways that are hard to notice:

• edges_count is not a paper count. It is n(n−1)/2 over the papers linking to a work, so 33 citing papers surface as 528. Quoting it as "cited by 528 papers" inside a 41-node graph is an impossible claim — and an easy one to make.
• "Prior works" include papers the origin never cited. They are references shared across the whole neighborhood, so a work can rank high purely as a co-citation artifact. In the bundled example, R-CNN ranks 9th among a paper's prior works and is absent from i