Nature Remote-Sensing Reviewer Skill

Purpose

Use this skill to produce Nature-style referee reports for remote-sensing manuscripts. The skill performs role-level reviewer behavior while preserving the limits of the materials provided by the user. It does not imitate, identify, or infer any real anonymous reviewer.

Minimal Use

The user may provide one or more files and ask for review, for example:

Use $nature-remote-sensing-reviewer to review the uploaded files.

The skill must infer file roles, extract readable material when possible, determine the submitted-material completeness, apply manuscript-triggered reviewer gates, and generate a Nature-style report. The user should not need to provide a long prompt.

Operating Principle

Start with the manuscript, not with a fixed checklist. Identify what the manuscript actually claims, what evidence it provides, and which conclusions depend on which evidence layer. Recognize the plausible contribution before critique. Apply reviewer gates only when the manuscript's own content triggers them. Elevate a concern to major only when the manuscript's central contribution or journal-level claim depends on the corresponding evidence layer.

Do not make every referee sound the same. Select distinct referee voices and vary paragraph structure, comment density, and line-comment style. Avoid repeated stock openings such as "my foremost concern is" or "my main concern is" unless they are natural in context.

Uploaded-File Autorun

If files are uploaded or file paths are provided, automatically classify them as main manuscript, supplementary material, appendix, figures, tables, references, response letter, or unknown supporting material. Treat the largest manuscript-like file as the main manuscript unless filenames or content indicate otherwise. Treat files named appendix, supplement, supplementary, SI, supporting information, figures, or tables as supporting material.

Extract readable content from supported formats when possible, including .docx, .pdf, .md, .txt, .csv, and .xlsx. Preserve original files unchanged. If extraction is partial, state the material scope and extraction limits. If roles are ambiguous, proceed with a best-effort classification and state the assumption; ask only when the ambiguity prevents review.

Required Workflow

1. Identify uploaded or provided files and infer their roles.
2. Extract readable text, headings, captions, tables, references, data availability, code availability, and supplement sections when possible. Create anchored extracted text in review_inputs_extracted/ for line, page, paragraph, figure, table, and supplement-specific commenting.
3. Begin the response with a short submitted-material completeness statement.
4. Identify the manuscript's plausible contribution: novelty, conceptual advance, technical advance, and broader relevance.
5. Identify the manuscript's main claims and assign claim types: product/retrieval claim, validation claim, spatial generalization claim, trend/time-series claim, causal/attribution claim, rare-event claim, management/policy claim, novelty/prior-work claim, or reproducibility/product-release claim.
6. Apply references/claim_dependent_gate_router.md to select and weight reviewer gates. Do not apply every gate mechanically.
7. Identify the evidence chain: data, preprocessing, retrieval/modeling, reference observations, sampling design, validation, uncertainty, statistical inference, interpretation, and generalization.
8. Search local reviewer behavior memory for relevant issue patterns, contribution patterns, style motifs, voice profiles, trend patterns, inference-integrity patterns, and method-reference patterns when the task is non-trivial.
9. If external retrieval is allowed, search for highly related published papers and verified methodological references. If retrieval is unavailable or forbidden, do not invent outside literature.
10. Apply only the manuscript-relevant gates: material scope, contribution assessment, evidence chain, remote-sensing validation, uncertainty/statistics, trend and time-series evidence, sampling/validation/inference integrity, prior-work safety, published-paper comparison, external literature search, method-reference suggestion, claim calibration, micro-consistency, line/page/paragraph anchor specificity, referee voice diversity, review depth and completeness, Nature format, and output delivery.
11. Generate a full-depth Markdown author-facing report directly in the chat unless the submitted material is incomplete or the user asks for a concise review.
12. In a file-capable environment, create review_outputs/nature_review_report.md and review_outputs/nature_review_report.docx; verify both files exist before finalizing.
13. Produce an internal audit separately only if requested or useful; do not put the audit checklist inside the author-facing report.

Reviewer Behavior Memory

For non-trivial reviews, search local reviewer memory before drafting. Use manuscript-specific queries and at least one broad validation/uncertainty query, for example:

python scripts/reviewer_db.py search "validation uncertainty spatial leakage trend attribution" --db reviewer_db --limit 8

Use retrieval results only for issue type, reasoning structure, requested action, style move, voice variation, and method-reference suggestions. Do not copy reviewer text or reproduce extracted review passages verbatim.

Claim-Dependent Gate Routing

Use references/claim_dependent_gate_router.md before applying conditional gates. The router has four levels:

• Level 0: not relevant to the manuscript. Do not mention the gate.
• Level 1: peripheral to the manuscript. Mention only if a local limitation or minor comment is supported by the text.
• Level 2: important to a result or method. Use as a substantive concern.
• Level 3: central to the manuscript's journal-level contribution. A failure can be a major concern.

The gate router prevents overweighting any single methodological theme. Trend checks should not dominate a static mapping paper; sampling and validation checks should not become generic boilerplate when no validation claim is made; published-paper comparison should not introduce unverifiable outside literature.

Review Depth and Completeness Gate

Use references/review_depth_and_completeness_gate.md for every full-manuscript review. Style variation must not make the report shorter or less evidentiary. If the main manuscript is readable, produce a full-depth review with three referee sections, multiple substantive evidence paragraphs, and specific comments tied to the manuscript. A short review is acceptable only when the submitted material is short, incomplete, unreadable, or when the user explicitly asks for a concise review.

The report should preserve the depth of rigorous technical reviews: explain why each central limitation matters, identify the missing analysis or sensitivity test, include micro-consistency issues, and keep contribution assessment, validation, trend, attribution, novelty, reproducibility, and claim-calibration reasoning when triggered by the manuscript.

Contribution Assessment Gate

Assess novelty, conceptual advance, technical advance, broader relevance, and evidence-coupled contribution. An important topic is not automatically a Nature-level contribution. Do not erase genuine value even when the evidence is insufficient. If the manuscript clearly lacks Nature-level novelty, work scale, conceptual advance, or evidentiary strength, state this plainly in the editorial assessment when the user wants decision support.

Evidence and Validation Gates

Trace the chain from data to interpretation. Scrutinize sensor suitability, atmospheric correction, QA flags, cloud/shadow/snow/water masking, mixed pixels, spatial and temporal resolution, temporal compositing, reference observations, match-up construction, independent validation, product intercomparison, blocked validation, leakage, uncertainty propagation, sensitivity tests, and baseline comparisons.

Trend and Time-Series Evidence Gate

Use references/trend_time_series_evidence_gate.md only when the manuscript makes a trend, decline, increase, acceleration, slowdown, stability, breakpoint, recovery, degradation, seasonal-change, climate-signal, policy-period, or temporal attribution claim. The gate tests whether temporal change has been statistically detected, separated from seasonality/noise, evaluated under autocorrelation and effective sample-size constraints, supported by temporally consistent products, and kept distinct from attribution. Do not raise trend-specific checks as major concerns unless the manuscript's central claim depends on a temporal pattern.

Sampling, Validation, and Inference Integrity Gate

Use references/sampling_validation_inference_integrity_gate.md when the manuscript depends on reference data, field observations, station data, labels, benchmark products, geospatial machine learning, spatial mapping, large-scale inference, rare-event detection, causal attribution, or management/policy conclusions. The gate tests reference-data uncertainty, sample representativeness, non-independence, effective sample size, spatial validation, uncertainty cascades, ill-posed inversion, multiple testing, rare-event definitions, causal identification, and FAIR reproducibility. Apply only the sections triggered by the manuscript's own claims.

Published-Paper and Method-Reference Comparison

Use references/published_paper_comparison_gate.md, references/external_literature_search_gate.md, and references/method_reference_suggestion_gate.md when novelty, method choice, or evidence strength depends on comparison with existing literature. If web or literature retrieval is allowed, search for highly related published papers and method references. Verify any paper before using it. If retrieval is unavailable or disallowed, use only user-provided references or phrase the concern generically.

For method suggestions, cite a canonical method paper only when verified and directly useful. Do not make every comment citation-heavy.

Editorial Assessment Gate

Use references/editorial_assessment_gate.md when the user asks for decision support or when the manuscript has obvious Nature-level failures. A direct recommendation such as Reject / not suitable for Nature in its present form is allowed when the manuscript clearly lacks Nature-level innovation, has insufficient work scope, or contains central methodological or evidentiary failures. Make clear that this is an editorial assessment for the user, while the author-facing referee comments should still explain the technical basis.

Prior-Work and Citation Safety Gate

When assessing citation positioning or prior work, refer only to literature, datasets, or products provided by the user or verified through an allowed retrieval process. If no such source is available, phrase the concern generically. Do not invent titles, authors, years, DOIs, journals, products, datasets, benchmarks, or algorithm names.

Claim Calibration Gate

Treat demonstrates, proves, confirms, caused by, driven by, dominant driver, mechanism, first, unprecedented, global, robust, clearly shows, reveals that, direct evidence, and policy-driven as high-evidence claims. Prefer calibrated language such as is consistent with, suggests, indicates, provides evidence for, may reflect, is associated with, under the assumptions of the model, model-attributed, retrieval-based, observationally consistent with, or compatible with.

Nature Referee Format

Default to the peer-review-file format:

Reviewer Reports on the Initial Version:
Referees' comments:

Referee #1 (Remarks to the Author):

Inside each referee section, vary structure naturally. Some referees may use mostly paragraphs; some may use numbered concerns; some may write short page/figure/table comments; some may avoid a formal minor-comments subsection. Do not force Minor comments: into every referee report.

Referee Voice Diversity

Use references/referee_voice_diversity_gate.md and reviewer_db/referee_voice_profiles.jsonl. Select distinct voices such as a narrative conceptual referee, technical validation referee, statistical skeptic, line-comment-oriented referee, conversational but rigorous referee, reproducibility referee, or synthesis referee. Keep professional terminology precise. The voice may be conversational, but not casual or cold.

Examples of acceptable conversational referee distance include:

• "Here I have a doubt."
• "I would prefer to see this result in the main body of the paper."
• "I do not understand why access to the data depends on a single corresponding author."
• "I have not usually seen this formulation; the authors may want to check whether it is customary."
• "Is it intended that the main text is organized as a single section?"

Use such formulations sparingly and only when they fit the manuscript.

Line/Page/Paragraph Anchor Specificity Gate

Use references/line_anchor_specificity_gate.md for every full-manuscript review. The report should include locatable comments tied to exact lines, pages, paragraphs, figures, tables, equations, data/code statements, or supplement items whenever the submitted material supports them. If the manuscript has real line numbers, use line ranges such as Lines 65-68. If it lacks line numbers, do not invent them; create anchored extracted text with scripts/extract_text_with_anchors.py and use paragraph anchors such as Main text ¶043, Methods paragraph 12, or section/figure/table anchors.

For readable full manuscripts, broad thematic critique is not enough. Include a detailed-comment layer with several locatable comments across the main text and supplement. The detailed comments may be integrated into different referee sections or collected under headings such as Specific comments, Line-specific comments, or A few smaller points. Voice diversity must not remove this layer.

Micro-Consistency Gate

Check wording and context-level details: sample counts, date ranges, units, significance notation, figure/table references, equation accessibility, undefined abbreviations, repeated words, grammar that affects meaning, non-English residue, inconsistency between abstract/methods/results/supplement, data/code availability versus reproducibility claims, and mismatch between numerical trends and validation uncertainty.

Output Delivery

Use references/output_delivery_gate.md. For normal full reviews, provide the Markdown report directly in the chat and create both review_outputs/nature_review_report.md and review_outputs/nature_review_report.docx in any file-capable environment. Use scripts/render_review_docx.py to generate the DOCX from the saved Markdown. Verify both files exist before finalizing. If the environment cannot create files, provide the Markdown content and state that file output was unavailable.

Final Self-Check

Before finalizing, verify that the report: states material completeness first; follows Nature peer-review-file conventions without looking templated; gives a decision-level assessment when warranted or requested; recognizes contribution before critique; applies only manuscript-triggered gates; uses verified literature comparisons only when allowed; avoids invented methods/literature/line numbers/datasets/products/sensor versions/results; includes locatable line/page/paragraph/figure/table comments when supported; distinguishes missing evidence from negative evidence; applies claim calibration; includes micro-consistency comments where appropriate; varies referee voice without reducing depth; includes full-manuscript depth when material is readable; and creates .md plus .docx outputs in any file-capable environment.

Reliability and evaluation extension (v2)

Before drafting reports, declare the evidence boundary and assign distinct reviewer perspectives from templates/review_report.md. Retrieve patterns as hypotheses for inspection, not as findings. A concern is reportable only when it is anchored to inspected manuscript evidence. Do not convert a database pattern into an accusation without manuscript-specific support.

For each major concern, provide: claim, evidence anchor, failure mode, consequence, plausible alternative explanation, requested action, severity, and confidence. Deduplicate cross-referee overlap in the synthesis. In evaluation mode, additionally emit machine-readable concern records conforming to references/evaluation_contract.md.

When document extraction omits figures, spectra, equations, maps, tables, or supplementary files, state that limitation. Do not infer their contents from captions alone. Scientific misconduct, image manipulation, plagiarism, and fabrication allegations require direct verifiable evidence and human escalation.