literature-triage-matrix

literature-triage-matrix

Produce a single comparison table over a set of papers, optimized for review-writing decisions. The matrix lets the user (or a downstream writing skill) see at a glance which papers cluster together by method, which carry the load on a particular claim, and which can be cited in passing vs deeply engaged.

This skill avoids the common AI failure mode of "give me a summary of each paper" — which produces N independent bullet lists that the user then has to manually compare.

When to use

Trigger phrases:

• "Make a literature matrix for these papers."
• "Compare these papers by method, data, claims, and limitations."
• "Help me decide which papers are central to my review."
• "Build a triage table over my Zotero collection / Obsidian cluster."

Not for:

• A 5-page narrative literature review — that's a writing task.
• Citation formatting — Zotero and the writing skill handle that.
• Single-paper deep dive — use paper-memory-builder.

Inputs

In priority order (cheapest to most expensive):

1. Manual paper list — a Markdown bullet list of titles + DOIs (or arXiv IDs) the user pastes directly into the chat. Lowest-friction entry; works without any other research-hub setup. Treat each line as one row in the matrix; fill cells from your own knowledge

   • DOI lookup if the title is famous, otherwise mark ? and ask the user.

   Example minimal input:

   - "Memory enables ToM-like behaviour in LLM poker agents", arXiv:2604.04157
   - "Multi-agent LLM social learning", arXiv:2604.02677
   - "Triadic Loop alignment framework", arXiv:2604.18850
   

2. .research/literature_matrix.md — if it already exists, parse it first. Append-only by convention; only re-emit a row if the underlying paper changed materially.

3. Obsidian cluster notes under raw/<cluster>/*.md — these have structured frontmatter (title, authors, year, doi) plus research-hub-generated Summary / Key Findings / Methodology / Relevance sections. Read these first; they're cheaper than PDFs.

4. Zotero collection metadata via local API (fast) — add child note contents only if Obsidian doesn't have the paper.

5. NotebookLM downloaded briefs under .research_hub/artifacts/ — if the user has already generated a brief on the cluster, mine it for cross-paper comparisons.

6. Raw PDFs — only as last resort, and only the abstract + first 2 pages + conclusion. PDFs are token-expensive.