Zotero Compare

If the comparison dimension is vague, ask — a focused comparison is far more useful than "compare everything." Don't extract content until the dimension is nailed down.

Step 2: Parallel triage — abstracts + outlines

For each item, get abstract and outline in parallel. These are all metadata operations (no PDF downloads), so issue them as one batch of parallel tool calls:

# All in parallel (one turn):
uv run zotero_read.py abstract KEY1
uv run zotero_read.py abstract KEY2
uv run zotero_read.py abstract KEY3
uv run zotero_read.py outline KEY1
uv run zotero_read.py outline KEY2
uv run zotero_read.py outline KEY3

Note: outline does download the PDF on first call (for the first run of that item), so it's not strictly free — but the PDF gets cached and is reused by the subsequent pages calls in Step 3.

From the abstracts and outlines, identify for each item the 1–2 sections most relevant to the comparison dimension. Write down the page ranges.

Step 3: Targeted extraction — parallel, budgeted

Extract the relevant page ranges for each item in parallel. Budget: ~15 pages per item, ~40 pages total max. This keeps token usage reasonable and focused.

# All in parallel (one turn):
uv run zotero_read.py pages KEY1 12-17
uv run zotero_read.py pages KEY2 8-15
uv run zotero_read.py pages KEY3 20-27

For short papers (fewer than ~20 pages in total), prefer fulltext — one call instead of a targeted range:

uv run zotero_read.py fulltext KEY_SHORT_PAPER

Mix and match: some items may use pages, others may use fulltext, depending on length.

Step 4: Synthesize the comparison

Pick the format that fits the comparison:

Table format (feature/property comparison)

Narrative format (nuanced conceptual comparison)

On time integration: Hesthaven & Warburton (2008, Section 4.3, pp. 89–95) use explicit RK methods with a CFL constraint proportional to element size over polynomial degree squared. By contrast, Ferrer (2012, Section 3.1, pp. 12–14) employs an implicit-explicit splitting that treats the stiff viscous terms implicitly, lifting the CFL restriction.

Both (for complex comparisons)

Summary table first, then narrative on the most interesting differences. Let the table carry the "what" and the narrative carry the "why it matters."

Rules

• Separate quotes from interpretation. Direct quotes in blockquotes with exact page refs, your comparative analysis clearly labeled. Same rules as zotero-read.
• Never fabricate quotes. If you can't find exact wording, describe what the section covers and cite the location: "Section 3.1 (pp. 12–14) discusses the IMEX splitting."
• Every claim must have a reference — (Author Year, Section X.Y, p. N).
• Be honest about gaps. If one paper doesn't address a comparison dimension, say so. A gap is informative.
• Highlight agreements AND disagreements. Where papers align is as useful as where they diverge.
• Stay focused on the comparison dimension. Don't produce a full summary of each paper — that's a different task.

Strategy tips

• Sequence: triage → targeted extraction → synthesis. Don't dump all pages of all papers into context upfront. Use outlines and abstracts to decide where to look before downloading content.
• Annotations as a shortcut. If the user has annotated any of the papers on the comparison dimension, annotations KEY gives you their highlights directly — use those first.
• Parallelism matters. Each zotero_read.py invocation is independent. Issue them as parallel tool calls, not sequentially.
• Budget discipline. If a paper's relevant section is 30+ pages, pick the core subsection and note "Pages 20–25 selected from a larger Section 4 (pp. 15–42)" so the user knows you narrowed it.