ingest-source

A Claude Code skill that imports external academic material — a paper, PDF, slide deck (PDF/PPTX), DOI, arXiv id, a URL to any of those, or pasted text — into a Mneme knowledge graph as linked markdown dots, clearly marked as imported (their own Lit:<Topic> lane + a source: field) so outside knowledge stays visually distinct from your own work.

Sibling of distill-session: that one captures your Claude Code sessions; this one captures what you read.

How it works

1. extract.py turns the source into plain text with local CLIs — pdftotext, pandoc, curl, and stdlib (a .pptx is a zip of XML, parsed directly; a DOI is resolved via the Crossref API). The model spends ~no tokens on the bytes.
2. A subagent reads the extracted text (outline first) and distils it into dots.
3. affiliate.py (from distill-session) places each dot by keyword overlap and suggests which of your own dots it should link to.

See SKILL.md for the full workflow.

Requires

pdftotext (poppler), pandoc, curl, python3 (stdlib only). Legacy .ppt needs LibreOffice; convert to .pptx/.pdf first.