rlm-research-analyzer

RLM Research Analyzer

A Claude Code skill that synthesizes a Zotero paper collection into a structured literature review — per-paper summaries, cross-paper themes, contradictions, gaps, and a research questions status table — delivered as Markdown, PDF, and Zotero notes.

The RLM pattern

This plugin is loosely based on the Recursive Language Model (RLM) inference pattern. RLM was designed to process documents too large for a single context window without RAG infrastructure: a root controller (frontier model) never reads the full data directly — it only sees metadata — and spawns cheaper worker sub-models to read slices, which return structured findings the root then aggregates. A validate pass checks the result for errors and inconsistencies.

This plugin adapts that pattern to a collection of independent papers:

RLM concept	Plugin implementation
Root sees metadata only	Stage 1 fetches titles and abstracts only — no full text
Workers read data slices	Stage 3 spawns one Haiku subagent per batch of ~5 papers; each reads full text via zotero_get_item_fulltext
Cheaper worker models	Workers run on Haiku; the root and reviewer run on Sonnet
Structured findings returned	Each worker writes a slice_N.md with standardized headings, confidence tags, and per-question verdicts
Root aggregates	Stage 4 consolidates all slice files into themes, agreements, and contradictions
Validate pass	Stage 5 runs an adversarial reviewer subagent twice, fixing all FATAL/MAJOR issues before delivery

The filesystem (slice_N.md files) serves as the persistent shared state between workers and the root — the equivalent of the REPL environment in the original pattern. No single model ever holds more than ~5 papers of full text at once, so the pipeline scales to arbitrarily large collections without context window saturation.

What it does

Given a Zotero collection name (and an optional focus question), the skill:

1. Indexes the collection — fetches metadata, runs semantic filtering if a focus question is provided
2. Plans — derives research questions, groups papers into batches, waits for your approval
3. Delegates — spawns parallel Haiku workers to read full text and produce structured per-paper summaries
4. Aggregates — consolidates findings into a synthesis with themes, agreements, contradictions, and gaps
5. Verifies — runs two adversarial reviewer passes; fixes all FATAL/MAJOR issues before delivery
6. Delivers — writes a final Markdown answer, a styled PDF, a provenance file, and creates a Zotero note on each paper with its summary

Prerequisites

Requirement	Role	Required?
Claude Code	Runs the skill	Required
Zotero desktop app (running locally)	Source library	Required
zotero-mcp MCP server	Provides all zotero_* tools	Required
GEMINI_API_KEY env var (or OpenAI equivalent)	Semantic search filtering when a focus question is provided	Optional
Python packages: weasyprint, markdown	PDF generation	Optional

Install optional PDF dependencies:

pip install weasyprint markdown

How it works

RLM runs in six sequential stages, each building on the last.

Stage 1 — Initialize

Claude fetches all item metadata (title, abstract, authors, year) from the Zotero collection via zotero-mcp — no full text at this point. If a focus question was provided, it runs a semantic search and intersects the results with the collection to narrow the working set. The full index is written to rlm_index_<slug>.md.

Stage 2 — Plan

Claude reads the index, infers 2–5 specific research questions from the abstracts, and presents a research scope for your approval. You can adjust the questions, include/exclude items, or tweak the anticipated sections. Once approved, items are grouped into batches of ~5 and the plan is written to rlm_plan_<slug>.md.

Stage 3 — Delegate

One Haiku subagent per batch is spawned in parallel. Each worker calls zotero_get_item_fulltext for every paper in its batch, falls back to the abstract if full text is unavailable, and writes a structured per-paper summary with:

• Key findings tagged with a confidence level (high = read in full text, medium = abstract only, low = inferred)
• Methods, conclusions, and coverage notes
• A verdict per research question (answered / partially answered / unresolved / superseded)

Results land in slice_N.md files.

Stage 4 — Aggregate