- name
SKILL.md
スキルファイルは公開元のリポジトリに置かれたままです。
コピーしたコマンドをターミナルに貼り付けてください。この SKILL.md がそのエージェントにインストールされます。
npx skills add 917Dhj/DeepPaperNote --skill deeppapernote -a cursor -yコピーしたコマンドをターミナルに貼り付けてください。この SKILL.md がそのエージェントにインストールされます。
npx skills add 917Dhj/DeepPaperNote --skill deeppapernote -a claude-code -yコピーしたコマンドをターミナルに貼り付けてください。この SKILL.md がそのエージェントにインストールされます。
npx skills add 917Dhj/DeepPaperNote --skill deeppapernote -a codex -yコピーしたコマンドをターミナルに貼り付けてください。この SKILL.md がそのエージェントにインストールされます。
npx skills add 917Dhj/DeepPaperNote --skill deeppapernote -a grok -yコピーしたメッセージを Grok Bot のチャットで送信してください。Bot がこの SKILL.md を GitHub から取得し、すべての Bot が共有するスキルに追加します。
Install this skill: https://github.com/917Dhj/DeepPaperNote/tree/HEAD/skills/deeppapernote公開元より
Turn one paper into a replication-grade deep-reading note and save it to the workspace or Obsidian target of the resolved configuration. Each run handles one paper. A rewrite of the abstract is never a finished note.
The note reconstructs the paper's argument as one story that runs through its middle sections: the problem it takes on, how the work answers it, what the evidence shows, what the paper does not prove, and why it is worth keeping. The paper type sets the middle sections and the questions they must answer (references/paper-types.md).
Write as a senior colleague explaining the paper to a smart researcher from an adjacent field: someone who reads papers but does not know this paper's subfield, its jargon, or its baselines. Lead each paragraph with its point, explain a term the first time it appears, and keep the technical depth the paper type calls for: equations and implementation detail for method, system, and theory papers; design, measurement, and argument for the others.
These hold at every step:
*_raw_sections.jsonl and *_source_manifest.json are the canonical reading material. Finish only from inspected full-text evidence, never from the title, abstract, or headings alone.After Configuration Readiness, resolve one output_language (zh-CN or en) and pass it as --language to every script. source_manifest.language_hint describes the source text only and never selects the note profile.
The value travels through Save Target Admission → Figure Plan → Figure/Table Decisions → Synthesis Bundle → Note Plan → Grounding Lint → Final Note Lint → Final Quality Review → Final Readability Review → Formal Save. Every artifact records it as output_language, and each script stops on a missing or mismatched value. Final Note Lint records note_sha256, so any edit after lint means rerunning it. Formal Save checks every binding before any save side effect. references/output-language.md describes the two profiles.
Run the steps in order. A step is complete at its done when, never because a partial artifact exists. At any point, scripts/run_state.py status --workdir <run_directory> --prefix <run_prefix> reports verified stages, stale artifacts, and the next_action; references/run-state.md explains how to bind the artifacts you write.
Configuration Readiness. Resolve Run Overrides from the explicit request, CLI, and process environment. Read User Configuration only for fields they leave unset; references/user-configuration.md covers admission, migration, repair, and Preference Changes. Done when the save mode, its target, and the output language are resolved and valid.
Acquire. scripts/run_pipeline.py acquire --input <paper> --prefix <run_prefix> resolves the identity, collects metadata, fetches the PDF, and extracts the raw source text, following Sources below. Without --workdir it picks a directory under ~/.deeppapernote/runs/ and prints workdir=; reuse that value in every later step, or pass a writable --workdir when a sandbox blocks the default. Done when the source manifest exists and is not blocked.
Save Target Admission. In Obsidian mode, run scripts/write_obsidian_note.py --preflight with the resolved title, --language, Vault, --source-manifest, and --output <run_directory>/<run_prefix>_admission.json. Its result is authoritative:
Both lint artifacts bind the plan through note_plan_sha256, and Grounding Lint also binds the source manifest and decisions. These are exact-byte hashes, so reformatting JSON invalidates them: rerun both linters after a plan change, Grounding Lint after a source or decision change, and Final Note Lint after a note edit.
Running scripts/run_pipeline.py without a phase performs both phases in workspace mode only; Obsidian mode needs Save Target Admission between them.
A normal request finishes in one run: text, figure decisions, image materialization, and save. Produce a text-only note only when the user asks for one.
Prefer the strongest source in this order:
acquire checks the desktop library through the Zotero Local API first. Its --zotero-mode is auto by default and falls back to the web providers when Zotero has no match; off makes no Local API request, and required demands a Zotero match. An explicit Zotero key must resolve locally. A trusted JSON artifact or explicit local PDF stays authoritative and skips this lookup. A confident local-library hit outranks any title-only web match. When a library match exposes only an attachment key and filename, scripts/locate_zotero_attachment.py finds the PDF under Zotero storage. In Obsidian mode, a matching local archive is verified and reused before any download; references/paper-archive.md covers version and source selection and Connector handoff.
Stop, report the blocked stage and the remaining stages, and ask for what would unblock it when:
pdf_url, direct PDF URL, arXiv or open-access sources, an accessible publisher PDF, DOI enrichment, and any other implemented fetch pathblocked with blocked_code: unreadable_text: ask for an OCR version or another copy, since rerunning cannot fix itWhen any other step fails, retry it or take a fallback this skill names. Otherwise stop as above. An abstract-only, provisional, or partial note is never the finished output.
# title heading. Include tags with exactly one papers/<domain> value, the admitted domain, and useful aliases; add date, doi, or arxiv_id when known, and omit unknown fields. Always include the fields of writing_contract.frontmatter_contract: paper_type matching the plan, methods, datasets, and tasks as lists of at most 5 central items, and read_date as YYYY-MM-DD. A list value that names a term in the bundle's glossary_terms is written as that term's link, copied verbatim; every other value is plain text. Formal Save refuses a frontmatter link that is not a glossary term note.#, ##, and ### headings. The top-level sections are exactly writing_contract.section_order: the fixed head, the middle sections of contracts_by_paper_type[paper_type], then the fixed tail, in that order. references/paper-types.md explains how to choose the type, answer its core questions, and write the weak-evidence callout. A non-trivial paper gets several meaningful ### subsections in its middle sections.writing_contract.core_info_fields, in order. Analysis belongs in later sections.zh-CN mode, with no later evidence or judgment added.workflow gets that subsection in its host section: an overview chain, then a step table that mirrors note_plan.workflow (references/paper-types.md).storyline: each section picks up where the previous one left off, and a paragraph that serves no part of the story is cut. Within a section, write ordinary paragraphs, one point each, conclusion first and details after. Explain each term or acronym in one plain sentence the first time it appears, and keep an established English name rather than translating it word for word.references/final-writing.md and references/obsidian-format.md hold the full writing and Markdown rules.
Say each phrase only when its condition holds:
Lint is a floor, not polish: while the readability review is pending, say the draft passed script lint but has not finished its final language review. When a run stops early, name the current stage and the remaining stages.
Final user report:
summary of scripts/run_state.py status carries these fields.new_domain is true, say that this domain folder was created, give its new_domain_reason, and suggest the organize-wiki skill to add the domain to the user's convention.figure_dropsnote_sha256reader_research_direction came from research_memory, say in one line that the transferable ideas followed the user's research memory.glossary_followup.status is suggested, suggest running paper-glossary on this paper and give its source_manifest and note_path; when regenerated is true, add that term links the glossary had added to the replaced note return only after that rerun. For any other status, say nothing about the glossary.user_notes.status: preserved kept the user's My Notes; untouched found none to keep; heading_missing found no My Notes heading. Only when user_notes.legacy_format is true, add that the kept block predates the current note format and may be old model-written text the user can delete.replaced_images entry lists other notes in also_embedded_by, say in one line that the new crop of that figure also shows in those notes.SKILL.md and the synthesis bundle carry a normal run. Read a reference only when its row applies:
python3 is older, use another interpreter of 3.10 or newer, such as python3.12, for the whole run; when none exists, stop and report the version found.scripts/check_environment.py only when a concrete dependency or integration problem blocks a step, and explain the required dependencies, optional enhancements, and downgrade behavior directly.Generate a high-quality deep-reading note for a single paper and write it into an Obsidian-style vault. Use when the user gives a paper title, DOI, URL, arXiv ID, Zotero item, or local PDF and wants a
---name: deeppapernotedescription: Generate a high-quality deep-reading note for a single paper and write it into an Obsidian-style vault. Use…---# skills/deeppapernote/SKILL.md概要
Generate a high-quality deep-reading note for a single paper and write it into an Obsidian-style vault. Use when the user gives a paper title, DOI, URL, arXiv ID, Zotero item, or local PDF and wants a polished Markdown note with strong structure, evidence-based analysis, and figure placeholders.
よくある質問
公開元から skills/deeppapernote/SKILL.md をダウンロードするか、パスをコピーしてリポジトリを開いてください。PlugMCP はスキルを掲載するだけで、ファイルはホストしていません。
公開元は https://github.com/917Dhj/DeepPaperNote で管理しています。
Claude、Cursor、Codex、Grok Build は、公開元のリポジトリから追加した SKILL.md を読み込めます。Grok Bot は、チャットでスキルの GitHub フォルダーのリンクを送るとインストールします。
domain_required, new_domain_reason_required, or domain_name_invalid | Choose the domain from the receipt's domain_overview as references/domain-routing.md describes, then rerun with --domain, adding --new-domain-reason for a new domain. Carry both to Formal Save. |
| any other conflict | Report it and stop. |
Workspace mode runs the same preflight with --save-mode workspace only to choose the domain. Done when an admitted result with its domain is saved.
Prepare. scripts/run_pipeline.py prepare --workdir <run_directory> --prefix <run_prefix> --admission <admission.json> (workspace mode omits --admission) extracts structure and PDF assets, plans figures, builds the complete figure/table decision table, and writes the synthesis bundle, all offline. It refuses to replace decisions that hold Visual Review work unless the user accepts repeating those reviews and you pass --discard-reviews. Rerun a failed phase whole. Done when the bundle exists.
Visual Review and Figure/Table Decision Freeze. Record a verdict for every review_pending item with scripts/plan_figure_table_decisions.py review, send any Figure/Table the note needs into review with promote, and drop a usable candidate the note has no use for with a recorded reason, as references/figure-placement.md describes. Every decision must satisfy writing_contract.figure_table_contract. Done when scripts/plan_figure_table_decisions.py summary --decisions <decisions.json> lists no review_pending_labels; its table gives each insert's embed.
Note plan. Read the bundle and the raw sections, choose the paper type as references/paper-types.md describes, and fill every value of writing_contract.note_plan_contract.template to create a short JSON note_plan with the exact output_language. Set reader_research_direction to the research direction the request states, verbatim; when the request states none and research_memory.status is ok, copy research_memory.suggested_reader_research_direction; otherwise leave it empty. Write storyline as three to five sentences that tell this paper's own story the way the authors build it, from the problem through the idea and the evidence to what it means; it is the outline the middle sections follow, not text for the note. Fill key_prior_work with the three to five works this paper builds on, and put in connections only those connection_candidates from the bundle that truly relate to this paper, each with a relation key and a one-sentence reason. references/evidence-first.md shows a complete plan. Done when Grounding passes.
Grounding. scripts/lint_grounding.py --note-plan <note_plan> --source-manifest <manifest> --bundle-json <bundle> --figure-decisions <decisions.json> --language <language> --output <grounding.json>. Done when it reports no issue; keep its JSON for Formal Save.
Draft. Write the note from the plan, following Note Shape, as one Markdown file in UTF-8 without BOM and with LF line endings. Done when the complete draft is on disk.
Final Note Lint. scripts/lint_note.py --input <note> --plan-file <note_plan> --language <language> --output <lint.json>. Done when every reported passes_* gate is true. A failed style gate is normal for several rounds: fix the reported text and rerun. Each zh-CN mixed-language issue names the English it counted in english_run or english_words; translate that wording, or wrap a retained proper noun in backticks, keeping formulas inline and proper nouns intact. Escalate to the user only when the same failures repeat unchanged across several rounds.
Final Quality Review and Final Readability Review. Follow references/final-quality-review.md for the reviewer prompt, the recorded final_quality_review, the two-revision limit, and the final_readability_review receipt. A readability edit repeats Final Note Lint and a fresh quality review. Done when both reviews pass for the current note bytes.
Formal Save. Run scripts/write_obsidian_note.py with --content-file, --plan-file, --lint-json, --grounding-lint-json, --source-manifest, --figure-decisions, --final-quality-review, --final-readability-review, the admitted --domain and --new-domain-reason, and any approved overwrite hash, each naming the exact file the linters and reviewers checked. It repeats admission, refuses missing, failed, or stale gates, refuses unresolvable [[links]], and materializes each insert into the paper-local images/ directory. Done when the save succeeds.
formula_rules$...$$$...$$writing_contract.takeaways_contract, one or two items each, and each item is one or two sentences: reusable ideas that name the paper's mechanism, where it transfers, and its precondition, then follow-up questions or experiments specific to this paper, at least one of which tests the paper's own claims or evidence. With a non-empty reader_research_direction, connect at least one transferable idea to that direction whenever a credible transfer exists, start every item that connects to it with takeaways_contract.judgment_marker, and say plainly when no credible transfer exists.writing_contract.connections_contract: the in-vault subsection lists exactly the planned connections as [[target|Title]] entries with their relation labels, or its fixed empty statement; the key-prior-work subsection lists the planned key_prior_work as plain text. Never write the bibliography, and never write a [[link]] whose target is not a bundle candidate. Prose outside Connections carries no [[link]]; paper-glossary links terms in the body after saving. Formal Save refuses links it cannot find in the vault, and any link in workspace mode. A recommended next paper must already appear here. references/obsidian-format.md shows the format.writing_contract.my_notes_contract.guidance there. Formal Save keeps the user's own block from the note it replaces.