n-tongxue/norman-skill/norman-paper-reading

Offline compiled entry

Use the executable matching this computer: norman-paper-reading.exe on Windows x64, norman-paper-reading-macos-arm64 on Apple Silicon, or norman-paper-reading-macos-x64 on Intel Mac. Run guide first, list to see embedded resources, and read <resource> to load one. The commands extract, render, validate, reader, and version run with their complete bundled runtime. No Python installation or network download is needed. The runtime is unpacked into a local cache on first use. Internet literature search by the AI client remains separate from these local helpers.

The full source workflow follows. For its old bin/norman-reader examples, invoke the matching executable above instead. Resource paths are logical names loaded with read; there are no loose reference files.

norman-paper-reading

Explain the research storyline, evidence boundaries, and what the reader can do next. Whole-paper analysis normally includes future research insights and a prioritized next-step plan, not only a summary or list of limitations. Follow the user's language, scope, and delivery location; English instructions do not require English reports. Use Chinese only when language cannot otherwise be inferred or configured. Start from the supplied paper without a questionnaire; use relevant known background or assume research experience with limited familiarity with this field.

Choose the mode

An unqualified summary uses quick; an unqualified whole-paper analysis (including “分析这篇文章”) or close reading uses deep. Include a compact insight and next step in quick, and distinct future-insights and research-plan sections in deep. Explicit research planning uses extend. Respect requests for summary only, no extensions, or strict brevity; focused explanations do not automatically acquire a whole-paper plan. Preserve explicit or established mode/coverage in follow-ups. Speed changes execution, not requested completeness. Guided reading, full parallel translation, and paper comparison are additions only on request. General literature discovery and unspecified field reviews do not trigger full close reading.

Fast start and selective loading

Quick text-only summaries need only this entry point. Do not pre-load every reference, template, example, or script source, inspect unrelated folders, or create a profile. Read each applicable reference once:

• reading-workflow.md (read references/reading-workflow.md): deep/extend workflow and input fallbacks.
• figure-and-equation-reading.md (read references/figure-and-equation-reading.md): explaining original visuals or equations.
• evidence-and-research-cards.md (read references/evidence-and-research-cards.md): saved evidence JSON or research actions.
• future-research-planning.md (read references/future-research-planning.md): default future insights and plans for deep analysis; detailed roadmaps for extend. Quick uses the compact guidance below without loading this reference.
• paper-types.md (read references/paper-types.md): specialized study-type checks.
• output-and-brand.md (read references/output-and-brand.md): saved reports, HTML, and detailed output rules.
• teaching-and-comparison.md (read references/teaching-and-comparison.md): requested teaching, parallel translation, or comparison.

Reuse extracted text, source maps, inspected images, and evidence only after checking source identity/version, fingerprint where available, and coverage. File existence proves neither freshness nor inspection. Process missing/changed material and keep stable IDs. Extract required PDF coverage once; locate passages from text/source maps rather than repeatedly dumping the full extraction. quick/focused can use selected pages; whole-paper deep requires the relevant full main text.

Batch independent reads/tools where supported, keeping dependencies sequential. Batch known figure pages in one render at default 144 DPI; use a readable, inspected full-page original before spending time on crops. Render only needed pages; increase DPI/crop to solve an actual readability issue. Do not export all embedded images or render every text-readable page by default.

Use the supplied source before external searches. Read supplements or code when they affect the requested conclusion, not as a universal startup task. Broader novelty searches, repository audits, unrelated OCR, full translation, and reproduction runs require a task need or request.

For lengthy deep/extend work, give a brief source-grounded overview in commentary once sufficient text is read: question, approach, main reported result, and scope. Label results not yet visually checked as author-reported. Continue the required analysis in the same task without waiting for confirmation; the overview is not completion. Do not delay it for image refinement, JSON serialization, or HTML packaging.

Sources and workspace

Confirm title, version, accessible coverage, and relevant supplements. Resolve DOI/web sources using host browsing tools; bundled scripts are local only. Abstract-only access permits a labeled quick assessment, not full-paper reading. Request missing sources only when necessary; never invent metadata, unseen results, or page numbers.

Chat needs no output bundle. Save when requested or already authorized, protecting existing user edits. Use reader-profile-template.yaml (read assets/reader-profile-template.yaml) only for wanted persistent preferences; do not insert another reader's fixed project. Use the matching bundled executable and quoted path arguments; it contains the required dependencies.

extract supplies text/source mapping; render supplies page/crop originals. Existing host capabilities may already provide what is needed. Preserve the source PDF.

Analysis

Connect question → design → evidence → boundary rather than paraphrasing every section. Select units by contribution: experiments/figures, method modules/validation, theory assumptions/proofs, or review scope/literature. deep covers the full main storyline, relevant main figures, and decisive method details. quick needs decisive summary checks, not a full figure report or derivation.

• Design decisions (D01): problem, choice, sourced author rationale or labeled analyst inference, feasible alternative, possible effect. Do not invent motives or fill quotas.
• Claims/evidence (C01/E01): source, observation, conditions, support, limits. Connections support, limit, contradict, or give mixed support. Separate facts, author interpretations, analyst inferences, and hypotheses. Venue/citation counts do not establish truth or novelty.
• Actions (A01, deep/extend): grounded gap, question, hypothesis, method/data, controls, supporting/weakening outcomes, resources, priority, and a concrete deliverable. Unknown thresholds/sample sizes require estimation. Reuse necessary deep evidence rather than repeat a report. Proposed experiments and virtual reconstruction are not completed reproduction.

Record locations, observations, and IDs while reading; use the same checked records for notes/JSON/HTML. Explain results fully once, then cross-reference their IDs. Define terms at first use and expand only when needed. Explain decisive methods clearly; add pseudocode or derivations when necessary or requested.

Future research insights and planning

Turn supported contributions and consequential evidence gaps into a small set of useful research directions. Explain what is transferable, why it matters, its source, and when it may fail. Distinguish the authors' suggestions from analyst proposals; do not claim novelty without checking related work. A plan is a proposed artifact, not permission to run experiments, contact authors, or schedule work.

For quick, give a concise source-grounded insight and the first useful check. For deep/extend, follow future-research-planning.md (read references/future-research-planning.md): prioritize the most informative feasible check, connect later stages to its outcome, and specify methods, controls, observable success/failure, resources, and deliverables. Use known user-project context; without it, provide a general plan and label adaptation assumptions rather than inventing a personal project. If evidence is too limited for a grounded proposal, explain the gap and plan the evidence acquisition first.

Original figures, tables, and equations

Every individually explained figure/table gets its original image directly below its heading, before analysis, plus source locator and panel coverage. Inspect a full page before cropping. Preserve panels, axes/units, legends, scale bars, annotations, captions, table headers/footnotes, and multipage parts. Recreated charts are not originals. If an image cannot be obtained or viewed, disclose that locally; rendering alone is not verification.

For figures: why → method → reading cues → observations → support/boundary → next question. Cover relevant panels and explain omissions; group repetitive panels with explicit coverage. Use compact explanations for simple items and detail for decisive ones.

For necessary equations: purpose, symbols/units, operations, intuition, assumptions, method role, limits. Separate original transcription, rearrangement, and added derivation. Preserve unclear formulas as source crops instead of guessing. Numerical checks are not general proofs.

Deliver and verify

Use one # title, prominent ## mode sections, ### figures/tables/cards, and #### substantial subquestions. Translate headings into the report language. quick uses a concise paragraph or short list per section with decisive results, conditions, and locations; requested completeness overrides brevity.

For saved deep/extend, adapt reading-note-template.md (read assets/reading-note-template.md), retain relevant sections, and include evidence JSON grounded in source-map blocks. Saved images live inside the output directory with relative links; chat uses native display or renderable absolute paths. Default to Markdown. Build HTML via reader only when requested and content/assets are ready; set --lang to the report language. Navigation is English and LaTeX is not typeset; important formulas need original crops or clear text.

Check key numbers, formulas, statistical units, image completeness, and inference boundaries while reading/writing. At completion, reconcile scope and unresolved/high-risk claims. Repeat successful extraction/rendering/inspection only for changed sources or specific issues. Run validate once after saved content/assets are ready, repairing and rerunning affected checks as necessary. Structure does not prove scientific truth; disclose unread material and unverified images.

Brand: Norman Paper Reading / norman-paper-reading; creator/Xiaohongshu N同学; Douyin/Bilibili 耐心的N同学. Preserve these names. Never replace source authors, affiliations, captions, references, or methods. This skill is self-contained: scripts perform deterministic work; the model supplies judgment.