| name | review-paper |
| description | Read academic review papers (PDF) and produce structured, critical digests. Use when asked to read, digest, summarize, or analyze a review article, scoping review, systematic review, or meta-analysis. Handles PDF extraction, review-type triage, field mapping, evidence auditing, and digest generation focused on synthesis quality rather than method validation. |
Review Paper
Read a review paper โ extract โ classify the review type โ map the field โ audit the evidence base โ file a critical digest.
Pipeline
PDF โ pdf_split.py --compact (deterministic Python, ~1 sec)
โโโ Sectioned markdown (5-15k tokens)
โโโ Metadata (title, authors, DOI, journal, year, keywords)
โโโ Figures โ disk
Read the markdown โ classify review type โ produce digest (template below)
+ image tool for key summary figures/tables when needed
โ file digest to project digests/ folder
โ file to YAML database with structured claims / follow-up papers
Default rule: use our local deterministic extraction pipeline first (pdf_split.py โ read markdown directly). Do not call the built-in pdf tool by default for normal review reading. Only use the built-in PDF tool as an explicit fallback when our extractor fails, the PDF is malformed in a way that breaks extraction, or there is a specific reason to compare outputs.
Why: the local pipeline is cheaper, more deterministic, closer to source text, and avoids quality loss from intermediate summarization.
Step 1: Extract
Run the bundled extractor:
python3 scripts/pdf_split.py <pdf_path> --output-dir <project>/extractions/<paper_name> --compact
Flags:
--compact โ default for pipeline; drops references, acknowledgments, strips noise (~35-40% token savings)
--no-refs โ only drop references section
--no-figures โ skip figure extraction
--json โ also output JSON structure (for programmatic use)
Output: <stem>_compact.md (or _extracted.md without --compact), figures/ dir.
If extraction fails or PyMuPDF is missing: pip3 install PyMuPDF then retry.
Step 2: Read and Classify
Read the extracted markdown fully. This is the core intelligence step โ no intermediate LLM.
Do not add a built-in pdf tool pass as a routine cross-check. If you want a cross-check, there must be a concrete reason: extraction corruption, suspicious metadata, missing sections/figures, or an explicit comparison request from the user.
First classify the review type
Before digesting, decide what kind of review this is:
- Narrative review โ expert overview, broad synthesis, often no formal search protocol
- Scoping review โ maps the literature landscape and coverage boundaries
- Systematic review โ explicit search and inclusion/exclusion process
- Meta-analysis โ quantitative pooling of study results
- Perspective / opinion review โ selective synthesis used to argue a position
- Tutorial review โ pedagogical overview rather than evidence-weighted synthesis
Record both:
- Review type
- Classification confidence: high / medium / low
If type is ambiguous, say so explicitly.
Step 3: Read with the right objective
A method paper asks: does this method work, and where does it fail?
A review paper asks different questions:
- How does this paper carve up the field?
- What literature does it privilege or ignore?
- Is the synthesis balanced, current, and evidence-anchored?
- What map does it give that primary papers alone do not?
- What should we read next to verify or deepen the claims?
Reading Strategy by Review Type
Narrative reviews: treat them as arguments, not measurements. Look for authorial bias, missing counterevidence, and selective framing.
Scoping reviews: judge breadth and map quality. Do not confuse coverage with evidence strength.
Systematic reviews: inspect search strategy, inclusion criteria, risk-of-bias handling, and whether the conclusions match the actual evidence.
Meta-analyses: inspect heterogeneity, pooling assumptions, non-independence, subgroup fishing, and whether a pooled estimate is being oversold.
Perspective / opinion reviews: useful for ideas, dangerous for consensus. Separate provocation from established evidence.
Step 4: Digest
Produce a structured digest following the template. Save to: <project>/digests/<YYYY-MM-DD>_<first_author>_<short_title>.md
The digest should stay concise and decision-oriented. The point is not to restate the review chapter by chapter; the point is to capture the field map, the evidence stance, the likely distortions, and the reading list hidden inside the paper.
Digest Template
All fields marked [ALWAYS] are mandatory. [BEST-EFFORT] fields filled when possible, marked "unknown" when not.
# {Title}
## Metadata [ALWAYS]
- **Authors:**
- **Journal/Preprint:**
- **Year:**
- **DOI/URL:**
- **Read date:** {YYYY-MM-DD}
- **Project:** {project_name}
- **Paper type:** review
- **Review type:** narrative | scoping | systematic | meta-analysis | perspective | tutorial
- **Classification confidence:** high | medium | low
## TL;DR [ALWAYS]
{2-3 sentences. What field it reviews, whether the synthesis is useful, and whether it is trustworthy.}
## Scope [ALWAYS]
- **Covered topics:**
- **Excluded / under-covered topics:**
- **Time span of literature:**
- **Intended audience:**
## The Review's Frame [ALWAYS]
- **Central thesis:**
- **How the field is organized:**
- **Assumptions in that framing:**
## Claim Cards [ALWAYS]
{3-7 key claims max. Each claim must be a falsifiable or checkable sentence and must include:}
- **Claim:**
- **Evidence basis:** meta-analysis | systematic synthesis | scoping map | narrative synthesis | expert opinion
- **Support level:** strong | mixed | weak | speculative
- **Anchor citations / primary papers:**
## Evidence Base [ALWAYS]
- **Most-relied-on primary studies / groups:**
- **Balance across methods/labs:** balanced | skewed | heavily skewed
- **Contradictory evidence acknowledged?:** yes | partially | no
- **Search / inclusion method reported?:** yes | partially | no
## Critical Assessment [ALWAYS]
### Strengths
{What this review genuinely does well.}
### Weaknesses
{What it does poorly, overstates, or glosses over.}
### Biases / Distortions
{Lab bias, method evangelism, citation circle, temporal bias, publication bias, etc.}
### Failure Modes [ALWAYS]
- **Outdatedness risk:** {is the field moving too fast for this review to still be reliable?}
- **What could mislead a reader:**
- **What must be checked in the primary literature:**
## Trust Verdict [ALWAYS]
- **Overall trust level:** high | medium | low
- **Why:** {2-4 concrete reasons only}
## Relevance to Project [ALWAYS when project active]
- **Significance:** high | medium | low โ {why}
- **Useful concepts / taxonomy:**
- **Gaps it leaves:**
## Top 5 Next Papers to Read [ALWAYS]
{Pick five primary or otherwise decisive papers, not a vague bibliography dump. For each item include:}
1. **Paper:**
- **Role:** foundational | best evidence | contrarian | methodologically sharp | best next-step for project
- **Why this paper:**
- **What claim from the review it helps verify/challenge:**
- **Priority:** high | medium | low
2. **Paper:**
- **Role:**
- **Why this paper:**
- **What claim from the review it helps verify/challenge:**
- **Priority:**
3. **Paper:**
- **Role:**
- **Why this paper:**
- **What claim from the review it helps verify/challenge:**
- **Priority:**
4. **Paper:**
- **Role:**
- **Why this paper:**
- **What claim from the review it helps verify/challenge:**
- **Priority:**
5. **Paper:**
- **Role:**
- **Why this paper:**
- **What claim from the review it helps verify/challenge:**
- **Priority:**
## Connections [BEST-EFFORT]
- **Related to:** {other papers in project database}
- **Contradicts:** {if any}
- **Extends:** {if builds on prior work}
## Tags [ALWAYS]
{comma-separated keywords}
## Raw Notes [BEST-EFFORT]
{Key tables, review figures, notable omissions, useful quotations, and any follow-up leads worth preserving.}
Hard Anti-Fluff Checks
A review digest is not acceptable unless these conditions are met:
- At least 3 claim cards are present unless the review is extremely narrow.
- No major claim without evidence anchors. If the review makes a claim but gives no real support, say so.
- At least one explicit missing piece or bias must be named.
- Review type must shape trust. Do not summarize a narrative review as if it were systematic evidence.
- Outdatedness must be checked. If the review stops before a major shift in the field, warn clearly.
- Do not mistake popularity for evidence. A highly cited narrative can still be weak.
Common Failure Modes to Watch For
Bias / incentive failures
- Citation laundering: authoritative language built on weak primary evidence
- Search-free selection bias disguised as expertise
- Overreliance on one lab, method family, or author network
- Failure to discuss negative or contradictory evidence
- Review as soft advertisement for a favored technique
Outdatedness failures
- Literature window ends before major methodological or conceptual shifts
- Definitions or standards changed after publication
- Foundational cited papers later weakened, contradicted, or superseded
Meta-analysis failures
- Pooling non-comparable studies
- Heterogeneity acknowledged but ignored in conclusions
- Multiple subgroup analyses used to rescue a weak main result
- Non-independence or double-counting of evidence
Step 5: File to Database
After producing the digest, add the paper to the project's YAML database. Same filing rules as the paper-reader skill.
Directory Layout
Same as paper-reader โ see that skill for the full directory tree. Typical paths:
- Digest โ
digests/YYYY-MM-DD_<author>_<short>.md
- PDF โ
papers_incoming/<key>.pdf
- Extraction โ
extractions/<author_year_short>/
- YAML โ
database_<collection>.yaml
YAML Schema
Same as paper-reader, but for reviews include:
paper_type: review
tags: ["narrative review", "cryo-EM", ...]
Connection Rules
- Always use
key strings โ the exact key field of the target paper
- Cross-database references are OK
- Bidirectional when meaningful
- Never use IDs, titles, or partial strings โ only exact
key values
Filing Checklist
- Copy PDF to
papers_incoming/<key>.pdf
- Write digest to
digests/YYYY-MM-DD_<author>_<short>.md
- Add YAML entry with all required fields
- Update connections โ add the new paper's key to related existing entries' connection lists
- Verify โ run a quick check that all connection keys resolve to real entries
ID Assignment
IDs may have gaps. Always assign max(existing IDs across all collections) + 1. Never reuse IDs.
Review-Specific Handoff Notes
Minimum items to preserve from the digest into the YAML maria_opinion or tags:
- Review type (narrative / scoping / systematic / meta-analysis / perspective / tutorial)
- Trust verdict
- Top follow-up papers identified (add these as connections where they exist in DB)
Principles
- Treat reviews as lenses, not neutral truth.
- Separate what the review claims from what the evidence actually supports.
- A review with no explicit search method is not automatically bad, but it is less evidentially trustworthy.
- A scoping review maps a field; it does not necessarily weigh evidence.
- A narrative review can be insightful and still badly biased.
- Always extract the hidden reading list. The best output from a review is often the set of primary papers worth reading next.
- Be concise. The digest should help someone decide what this review is good for, what it gets wrong, and where to go next.