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.

Evidence contract

Read ../research-hub/references/research-protocol.md and ../research-hub/references/source-claim-audit.md for scope preservation, core/closest/classic roles, version-aware findings, and bounded completion. If a standalone install lacks these references, use the minimum rules below: metadata, abstracts, remembered results and generated summaries are screening leads; load-bearing claims need located source/version evidence; unchecked claims and failed searches remain explicit gaps.

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. DOI lookup establishes metadata, not findings. Remembered famous results are search leads only. Retrieve actual source passages for claim/method/limitation cells; otherwise mark ? or clearly label the provisional screening source.

   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 for screening; generated sections are not independently verified source findings.

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 provisional comparison leads, with generated-summary labels.

6. Actual source text / PDFs — read the abstract for screening, then relevant methods/results/limitations, supplements or corrections for load-bearing detail. Report actual sections read; a fixed first-pages window does not constitute full-paper review.

Output

Write to <project-root>/.research/literature_matrix.md. Append by default; rewrite the whole table only if the user explicitly says "regenerate from scratch".

Markdown table columns (customize per request, but default is):

| Citation | Question | Method | Data / study area | Main claim | Evidence / source level | Limitation | Relevance | Use as |
|---|---|---|---|---|---|---|---|---|
| Smith 2024 | How does adaptation affect flood risk? | mesa-based ABM, 10k agents | Houston synthetic | Adaptation cuts loss 18% | simulation | single basin, no validation | High — direct precedent | Lit review §2 |
| Jones 2023 | Hydraulic coupling in flood ABMs | Coupled ABM-2D | Galveston | Coupling reduces RMSE 22% | empirical | calibration window narrow | Medium — methods | Methods §3 |

Column meanings:

• Citation — short reference (e.g., Smith 2024); leaving full citation to Zotero.
• Question — the paper's research question, one sentence.
• Method — model class + key technical detail.
• Data / study area — datasets used + geographic / temporal scope.
• Main claim — decision-relevant finding at the strength the reviewed source supports; ? if unverified, never a remembered result alone.
• Evidence / source level — study type plus actual material reviewed: metadata / abstract / generated summary / located full-text excerpt / unavailable. Include version, passage locator and audit reference for load-bearing claims; do not confuse study type with evidence availability.
• Limitation — most important caveat, one sentence.
• Relevance — High / Medium / Low + one-phrase justification.
• Use as — where in the user's manuscript or review this paper fits (Lit review, Methods, Discussion, citation-only). For review-selection decisions, name core / closest / classic / method comparator / contrary role, need IDs, omission effect, alternatives and applicability conditions in a compact companion note rather than silently adding unwanted columns.

Token-saving behavior

• Read existing .research/literature_matrix.md first; only emit rows for papers not already covered or whose underlying note changed.
• Prefer frontmatter and summaries for candidate screening. Reuse a prior verified passage only if source bytes/version and dependent decisions are unchanged; stale summaries cannot establish current support.
• Use stable paper identifiers (DOI or arXiv ID) so future requests can reference rows by Smith 2024 (10.1234/abcd) instead of pasting paragraphs.
• Declare a proportionate read/resource bound. If three reads are the approved bound, a fourth load-bearing source remains explicitly unassessed or is checked after an authorized continuation; the cap never certifies complete verification. Count all assessed and unchecked claims.

Output format for the user

After writing the matrix, print:

[literature-triage-matrix]
  Wrote/appended: .research/literature_matrix.md
  Papers in matrix: 12 (4 new this run)
  Skipped (already in matrix, unchanged): 8
  Source sections read: 1 (Smith 2024 — Methods §3 / Results §4)
  Claims checked / total: 3 / 4 (one load-bearing claim unassessed)
  Completion: partial; fourth source remains unchecked
  Suggested next: review High-relevance rows in .research/literature_matrix.md

What NOT to do

• Don't write to .paper/ — that's paper-memory-builder's territory.
• Don't fabricate findings if a paper's note is sparse. Emit the row with ? in unknown columns and a note: (Obsidian note incomplete — run research-hub auto with --max-papers 1 to enrich, or fill manually).
• Don't re-summarize unchanged verified evidence unnecessarily. Re-check claims when a publication version, correction or dependent decision changes.
• Don't include columns the user didn't ask for. The default 9 columns are a starting point; trim if requested.