ingest-pdf

name	ingest-pdf
description	Ingest a local PDF (or a folder of PDFs) into the Obsidian vault -- extracts DOI/title, resolves metadata via Semantic Scholar, writes 30_Literature/CITEKEY.md.

Ingest PDF (Desktop)

Use this skill when the user has a PDF already on disk (downloaded by hand, emailed by a collaborator, scraped from a publisher) and wants it captured into their vault as a proper 30_Literature/CITEKEY.md note. This is the companion to paper-capture, which only handles DOI / arXiv id / URL / title-author-year inputs.

Inputs accepted

• A single file path: D:\downloads\smith2024.pdf, /mnt/d/papers/raw/foo.pdf, or any absolute path.
• A folder of PDFs: e.g. {vault}/80_Attachments/papers-inbox/. Process every *.pdf under it; do NOT recurse into subfolders.
• No path given: ask the user once for the path or folder, then resume.

Workflow

1. Confirm vault path. Read OBSIDIAN_VAULT_PATH from the env. If unset, ask the user once.
2. Enumerate PDFs. If the input is a directory, list its top-level *.pdf files. If it's a file, that's the list of one.
3. For each PDF, extract metadata in this order:
   • (a) DOI sniff from text. Use the Read tool to read the PDF (Claude Desktop's Read tool can ingest PDFs directly). Extract page-1 / page-2 text. Run regex \b10\.\d{4,9}/[-._;()/:a-zA-Z0-9]+\b to find DOIs. If exactly one match: that's the DOI.
   • (b) arXiv id sniff. If no DOI, look for arXiv:\d{4}\.\d{4,5} or https?://arxiv\.org/abs/\d{4}\.\d{4,5}.
   • (c) Title + first-author + year heuristic. If neither id is present, parse the first-page text per references/pdf_metadata_heuristics.md: title is usually the first non-trivial line, first-author is the first name in the author list, year is the four-digit number near the title or in the running header.
4. Resolve to canonical metadata. Hand the DOI / arXiv id / (title + author + year) to paper-capture. paper-capture does the actual Semantic Scholar lookup, citekey generation, de-dup against existing 30_Literature/, and note write. Do NOT re-implement that logic here.
5. Move + rename the PDF. Once paper-capture reports a citekey, move the original PDF to {vault}/80_Attachments/papers/{citekey}.pdf. If the file already exists at the destination (different paper, citekey collision), defer to paper-capture's collision rule (suffix -b, -c, ...).
6. Update the note's frontmatter so pdf: points to the renamed attachment path. paper-capture writes pdf: null by default; ingest-pdf replaces null with the real path after the move.
7. Report. One line per ingested PDF: Ingested: {citekey} ({title trunc 60}) <- {original-path} Plus a final summary: Total: {n} ingested, {m} skipped, {k} failed.

Failure handling

• Encrypted / scanned-image-only PDF. No selectable text; DOI sniff returns nothing and the title heuristic is unreliable. Skip the file, report: Skipped: {path} -- no extractable text. Run OCR first.
• DOI doesn't resolve in Semantic Scholar. paper-capture will fall back to paper-mcp and then escalate to user. ingest-pdf surfaces that to the user and leaves the PDF in place.
• PDF is already in the vault ({vault}/80_Attachments/papers/{citekey}.pdf exists for the same DOI). Skip; report Already captured: {citekey}.

What this skill does NOT do

• It does NOT OCR. Image-only PDFs are reported, not handled.
• It does NOT re-implement metadata resolution. That's paper-capture's job.
• It does NOT write to 30_Literature/ directly. Always goes through paper-capture.
• It does NOT auto-tag the new note with a project. The user does that manually or via paper-map later.