菜单
Use when a quest has enough evidence to draft or refine a paper, report, or research summary without inventing missing support.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/ResearAI/DeepScientist --skill write命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Use for natural-science or engineering tasks, scientific software routing, simulation, dataset analysis, model fitting, package checks, HPC-through-shell work, validation, and evidence-backed scientific claims using DeepScientist's `artifact.science(...)` Science Evidence Graph. Includes a progressive-disclosure catalog of FermiLink skilled-scipkg package cards.
Prepare, audit, or revise Nature-ready Data Availability statements, data repository plans, dataset citations, and FAIR metadata checklists for manuscripts. Use when the user asks about Nature data availability, research data sharing, repository selection, accession numbers, restricted or sensitive data, source data, supplementary datasets, DataCite-style dataset references, FAIR metadata for academic publication, or Chinese-to-English data availability wording for Chinese-speaking authors preparing Nature-family submissions.
Submission-grade Nature/high-impact journal figure workflow for Python or R. Use whenever the user asks to create, revise, audit, or polish manuscript figures, multi-panel scientific plots, or journal-ready SVG/PDF/TIFF outputs, especially for Nature-family or other high-impact journals. Before plotting, define the figure's conclusion, evidence logic, export needs, and review risks. If the user has not chosen Python or R, ask "Python or R?" and stop. Use only the selected backend for figure generation, previewing, exporting, and QA. Supports matplotlib/seaborn and ggplot2/patchwork/ComplexHeatmap. Not for dashboards or Illustrator/Figma-first infographics.
基于 SOC 职业分类
正在显示 SKILL.md
| name | write |
| description | Use when a quest has enough evidence to draft or refine a paper, report, or research summary without inventing missing support. |
| skill_role | stage |
Draft To Top Conference Oral section below.Before editing a manuscript, first produce a concrete revision strategy from the current evidence state. Do not begin polishing prose until the strategy separates:
For each issue, choose exactly one action:
Never make an unsupported claim sound more convincing. If evidence is missing, either obtain evidence, narrow the claim, or mark the blocker.
memory.list_recent(scope='quest', limit=5) plus one writing-relevant memory.search(...). If restart context is unclear, use artifact.get_quest_state(detail='summary'), artifact.read_quest_documents(...), or artifact.get_conversation_context(...).paper/selected_outline.json, paper/evidence_ledger.json, and paper/paper_experiment_matrix.md or .json aligned. Use artifact.get_paper_contract(detail='full') as the default paper-reading surface when section rows, experiment rows, or analysis rows matter. Use artifact.get_paper_contract_health(detail='full') when outline state, experiment rows, or evidence ownership may be stale. Use artifact.submit_paper_outline(mode='candidate'|'select'|'revise', ...) instead of leaving outline choice only in prose.
When several paper shapes are plausible, record one or more outline candidates with artifact.submit_paper_outline(mode='candidate', ...), then select or revise explicitly with artifact.submit_paper_outline(mode='select'|'revise', ...); do not force extra outline rounds once the selected outline is good enough for the current writing job.artifact.validate_academic_outline(detail='full'). If it fails, use paper-outline or artifact.submit_paper_outline(mode='revise', ...) to repair the paper idea, claims, evidence boundaries, and analysis plan before prose work. When it passes, run artifact.compile_outline_to_writing_plan(detail='full') and draft from those jobs.breadth -> shortlist -> depth. Use DeepXiv or OpenAlex for discovery when available, then retrieve BibTeX from DOI or arXiv, not from memory. Keep paper/references.bib machine-usable and audit it before bundle submission.
If DeepXiv is declared available by the system prompt, prefer it for paper-centric discovery and shortlist triage before broad web search when it can answer the question directly. If DeepXiv is declared unavailable, do not try to force it; stay on the legacy route. Use for actual arXiv paper reads before escalating to full text.Do not let structural readiness stand in for paper quality.
draft_checkpoint_ready mean only that a package exists.paper-outline.artifact.validate_manuscript_coverage(detail='full').analysis-campaign, write an explicit analysis-budget waiver that downgrades the paper scope, or narrow the claims. Do not hide the shortage with prose.decision for a recommended stop or branch; record any narrowed non-paper objective as the next direction. If the recommended action is stop because paper quality is too low, ask the user to confirm before ending the paper objective. Consider user publication, scope, cost, or non-paper preferences before routing, and ask when the preference would change the route. Do not use polished prose to keep an unpublishable paper line alive.artifact.get_paper_contract_health(detail='full'):
use when a weak section may actually be caused by stale outline state, unresolved experiment rows, or unclear evidence ownership.artifact.get_paper_contract(detail='full'):
use by default before drafting any section, table, or analysis prose that depends on concrete main-experiment rows, analysis rows, or section-level result_table content.artifact.validate_manuscript_coverage(detail='full'):
use before bundle submission or finalize; it checks sections, displays, ready analysis groups, PDF, and checklist state.artifact.validate_academic_outline(detail='full'):
use before serious drafting; it checks whether the outline has a paper idea, scoped claims, evidence boundaries, method, evaluation plan, and enough planned analyses.artifact.compile_outline_to_writing_plan(detail='full'):
use after the outline is valid; it turns the outline into section-level writing jobs.artifact.validate_manuscript_language(detail='full'):
use after major prose edits and before submission; it catches route/user/worktree/port/batch wording that should not be in main text.artifact.get_quest_state(detail='summary'), artifact.read_quest_documents(...), artifact.get_conversation_context(...):
use when restart context is unclear, when exact durable wording matters, or when you need file truth instead of chat recollection.artifact.submit_paper_outline(mode='candidate'|'select'|'revise', ...):
use when outline choice or outline repair becomes durable enough that the paper line should follow it.artifact.create_analysis_campaign(...):
use only when a real paper-facing evidence gap needs follow-up analysis; do not use it for prose cleanup, citation chores, or generic "improve the paper" tasks.artifact.submit_paper_bundle(...):
use explicit package_type: draft_checkpoint, review_package, or submission_package only after coverage is submission-ready.artifact.interact(...) or other durable artifact updates:
use when the writing pass materially changes paper status, route choice, or bundle readiness and the change should survive beyond chat.bash_exec(...):
use for any real shell/CLI work such as LaTeX compile, bibliography checks, rg/find/ls, figure-generation scripts, PDF render/proofing, git inspection, or reproducibility checks. Do not describe command plans as if they ran; run them through when execution is actually needed.Follow the shared interaction contract injected by the system prompt. For ordinary active work, prefer a concise progress update once work has crossed roughly 6 tool calls with a human-meaningful delta, and do not drift beyond roughly 12 tool calls or about 8 minutes without a user-visible update.
paper_contract_health as a substitute for reading the actual section result_table, evidence rows, or experiment-matrix rows.write when paper-plot should own the first-pass figure.nature-polishing to make unsupported, stale, or overbroad claims sound stronger.nature-data to invent repositories, accession numbers, DOIs, licences, embargoes, access committees, or ethics approvals.nature-paper2ppt unless the user asked for an actual presentation deck.evidence_ready or analysis_ready as equivalent to manuscript_ready or submission_ready.selected_outline_ref / active paper line.decision instead of accumulating more draft text.the user requested, the latest user requirement, paper restart, this quest, the agent, the worktree, we were told, he accepted, paper should, or inside a paper draft.paper/selected_outline.json, paper/evidence_ledger.json, paper/paper_experiment_matrix.md or .json, paper/references.bib, paper/claim_evidence_map.json, paper/paper_bundle_manifest.json.bash_exec(...).artifact.create_analysis_campaign(...) only for real paper-facing evidence gaps, not for prose cleanup or citation chores.artifact.submit_paper_bundle(...) only after draft, bibliography, and bundle metadata are durable enough to hand off.analysis-campaign.memory.write(...) only for reusable writing, citation, or search lessons, not one-off local edits.30-50 verified references unless the scope clearly justifies fewer.paper/latex/ with a real template from templates/; for general ML or AI writing with no stronger venue constraint, default to templates/iclr2026/.paper/paper_experiment_matrix.md, , and / when relevant analysis results are meant to support the active paper line.result_table, evidence-ledger rows, or experiment-matrix rows rather than health-only summaries.paper/references.bib is real, current, and not hand-written from memory.bash_exec(...) execution rather than hypothetical prose.finalize-ready, currently feasible non-optional experiment rows are no longer unresolved.finalize-ready, artifact.validate_manuscript_coverage(detail='full') reports submission_ready=true; manuscript_ready=true alone routes to review, not finalize.Before writing or revising any paper-facing section, sort the source material:
Examples:
The nature-* skills are focused companion skills adapted from Yuan1z0825/nature-skills.
They can improve specific manuscript surfaces, but they do not replace DeepScientist's paper contract.
Use them as a short handoff inside the write flow:
artifact.get_paper_contract(detail='full') or the relevant quest documents for the evidence rows and missing fields that the surface may mention.nature-* skill and any referenced files it says are needed.write and update the durable paper surfaces before claiming progress: draft files, paper/evidence_ledger.*, paper/paper_experiment_matrix.*, paper/references.bib, figure/table catalogs, or bundle manifests as applicable.nature-polishing: use for Nature-leaning English, section restructuring, and Chinese-to-English academic polish. Apply it after the evidence boundary is clear, and keep unsupported claims downgraded or marked as blockers.nature-data: use for Data Availability, source-data, repository, dataset-citation, restricted-data, and FAIR metadata sections. Draft from verified inventory and leave unresolved fields explicit.nature-figure: use for Nature/high-impact-journal figure packages when figure claim, panel logic, backend choice, journal export, and QA are the main job. For simple structured result charts, prefer paper-plot first.nature-paper2ppt: use only for PPT/PPTX deliverables such as journal-club, lab-meeting, or paper-sharing decks. The expected output is a real deck plus lightweight verification.Routing examples:
nature-polishing, revise only the section job, then validate claim-evidence support.nature-data, inventory datasets and repositories, draft unresolved fields explicitly, then sync the section and references.nature-figure; if the job is only a simple result chart, stay with paper-plot plus figure-polish.nature-paper2ppt; keep it outside the manuscript bundle unless the user asks to attach it as a deliverable.references/oral_package_patterns.md when the draft needs a clearer oral-style evidence package.references/oral_writing_principles.md when the narrative spine, reader onboarding, or reviewer-facing tone is weak.references/experiments_analysis_patterns.md when experiments and analysis need clearer job separation.references/section_rewrite_checklist.md before treating a rewritten section as stable enough for bundling or review.Use this skill when a paper already exists in draft form and the real problem is not "write a paper from zero" but "turn this draft into something that reads like a top-conference oral paper."
This skill is for the transition:
Do not use this skill to invent missing evidence. If the draft has real evidence gaps, narrow claims or route to more experiments instead of hiding the weakness with better prose.
This skill is specifically about oral-paper upgrade work, not generic prose cleanup. It optimizes:
Read references/oral_package_patterns.md early when deciding what to add, cut, move, or split.
Use this skill when:
Do not use this skill when:
Read the current abstract, introduction, method, experiments, analysis, conclusion, and appendix if present.
Extract:
C1-C3: the 1 to 3 core claimsClassify the draft weakness into one or more of:
If the main issue is evidence, do not proceed as if this were only a writing problem.
Use references/oral_package_patterns.md to compare the current draft against an oral-ready target.
Label the biggest gaps. Typical gaps include:
When two versions of the paper exist, explicitly write the delta:
Top-conference oral papers are not just more polished. They spend pages and displays where reviewer friction is highest.
Before rewriting paragraphs, decide:
Default main-text priorities:
If the paper's central claim is comparative, benchmark-driven, or baseline-beating, the "main result display" must stay competitor-inclusive.
That usually means:
Do not collapse a broad benchmark story into a self-only summary table if the prose still makes broad comparative claims.
When the gold oral package keeps both a compact setup or baseline taxonomy and a competitor-inclusive benchmark surface in main text, preserve both jobs in the rewrite. Do not jump straight from prose setup to compressed averages if the reviewer still needs to see who was compared, under which regime, and where the main ranking or boundary actually appears.
When the paper has multiple proof obligations, do not present them as one continuous "results" stream.
Instead, turn the main empirical body into explicit reviewer-question blocks, where each block has:
If the strong paper or staged package already separates a section into named internal jobs, preserve that internal scaffold in the rewrite.
Do not collapse those jobs into one continuous wall of prose when reviewers need to inspect them separately.
This is especially important for:
When the paper's credibility depends on first proving that a metric, proxy, or diagnostic predicts reviewer-relevant outcomes, allocate a standalone validation block before intervention or design-guidance blocks.
Do not bury that proof inside later intervention subsections or leave analysis with only mechanism commentary if the draft package signals validation as the bridge into the rest of the paper.
If the draft package or staged artifacts separate several intervention families, keep them separate in the rewrite.
Each intervention family should still preserve:
If the evidence package carries multiple transfer fronts, keep at least one non-headline transfer benchmark or cross-setting validation in the main experiments section beyond the primary deployment or headline benchmark.
When the gold oral package uses multiple main-text displays to answer distinct reviewer questions, keep one explicit main-text boundary, robustness, or scope-setting display in addition to the headline comparison block. Do not push every non-headline empirical check into appendix overflow if the central claim still depends on visible claim-boundary evidence.
Only move exhaustive rows, per-task detail, and secondary checks to the appendix; do not narrow the main paper to one deployment table plus appendix overflow when the central claim depends on visible generalization breadth.
When the method makes a core claim operational, reserve method-local evidence for that claim.
For claims about open-ended actions, executable control, retrieval-grounding, tool use, or interaction loops, include at least one concrete method artifact when available:
Do not push all operational concreteness into experiments or appendix material.
Move exhaustive material to appendix:
Default appendix blueprint when the paper is mature enough:
Before drafting, record which main-text section must point to each appendix bucket.
Method, experiments, and analysis should each know which overflow material they are delegating and where the bridge sentence will appear.
Related work should also know whether it needs a bridge to an extended-literature appendix lane.
Generic appendix references are not enough when the manuscript relies on overflow evidence for credibility.
Each important bridge should name a precise appendix destination such as:
Do not write only "see the appendix" when the claim depends on protocol detail, method implementation detail, transfer overflow, extended literature, or worked traces.
When compressing a strong paper, do not let the appendix degrade into a light method bridge.
The appendix should still look like a reviewer-support package with explicit jobs, especially when the main text has compressed:
Top-conference oral papers stage information in the order that minimizes reviewer friction.
Rewrite in this order:
When writing the paper in a sectioned workflow, use this concrete generation order:
section_planintroductionrelated_workmethodexperimentsanalysisappendixlimitationsconclusionabstractintegrationUse section_plan as an internal control document, not as manuscript prose. It should record:
C1-C3Write the abstract last, after the paper's actual evidence order has stabilized.
In sectioned mode, keep main.tex as the canonical top-level document and keep body prose in separate section files. Do not collapse the manuscript back into one giant draft while writing. Use the final integration pass only to repair consistency, sharpen transitions, synchronize claim wording, and remove staging artifacts from the prose.
Do not reserve essential evidence allocation for integration. Each body section should already be locally complete enough that an interrupted integration pass does not erase key reviewer-defense blocks or appendix bridges.
Use the principles in references/oral_writing_principles.md.
The most important rules are:
When actively rewriting, use references/section_rewrite_checklist.md.
That file gives a practical pass for:
A mature oral paper does not merely mention likely reviewer concerns. It allocates explicit evidence to them.
Typical evidence blocks include:
If a likely objection matters, do not hide the answer in one sentence.
If the draft package supports several objection-resolving blocks, keep them as separate visible subsections rather than folding them into one omnibus paragraph or one overloaded table.
When the paper has enough evidence, reserve one explicit main-text block for reviewer-concern handling rather than hoping the reader infers those answers from the benchmark summary alone.
Typical reviewer-concern blocks to surface in the main text include:
For each evidence block, make the prose-display contract explicit:
If a section feels weak, diagnose the real cause:
Never use polished language to conceal an unaddressed scientific gap.
When the draft is dense enough to support staged writing, prefer generating the manuscript section by section rather than asking for the full paper in one turn.
Use these operating rules:
Introduction should not collapse a display-led first page into prose. When the staged package supports both problem scale and solution shape, preserve both roles with concrete displays, authored compact tables, or a figure-plus-table pairing.Introduction should preserve one concrete first-page failure case, benchmark contrast, or payoff anchor when the gold oral package uses it to make the problem vivid before formal sections begin.Related Work should name the closest prior and the exact novelty boundary rather than stopping at broad capability buckets.Method should keep a short main-text audit surface for model suites, benchmark groups, or regime inventory when the gold paper uses one to make the method's evidence base inspectable.Experiments should establish the main empirical pattern through explicit reviewer-question blocks, each anchored by one dominant display.Experiments should keep one non-headline transfer or robustness block in main text when the staged package has several transfer fronts and the central claim needs visible generalization breadth.Experiments should preserve visibly separate internal layers for headline evaluation, transfer breadth, and mechanism validation when the staged package distinguishes those jobs. Do not compress them into one undifferentiated benchmark narrative.Experiments should preserve repeated setup/results scaffolds for distinct intervention families when the gold oral paper uses them to turn validation into actionability. Do not collapse several intervention families into one short summary block if reviewers still need to inspect them separately.Method should preserve main-text setup and study-regime inventory when the draft package contains them. If the staged package distinguishes prediction settings, model suites, checkpoint slices, benchmark groups, or measurement definitions, keep those distinctions through separate subsections or strong subsection headings instead of pushing them all into appendix prose.Method should keep at least one local operational artifact when a core mechanism claim depends on concreteness, especially for executable action spaces, tool calls, browser actions, retrieval grounding, or closed-loop control.Method should preserve visible internal scaffold when the system explanation has distinct jobs such as workflow overview, specialist model design, supervision/data construction, and executable action realization. Strong paragraph heads are acceptable; one merged prose block is not.This audit should flag:
In this mode, a strong default main-text display program is:
If one of these roles is missing, do not merely mention it in prose. Either promote a staged artifact into that role or narrow the paper's claims to match the thinner package.
Before stopping, check:
A draft often tries to maximize information density. An oral paper maximizes comprehension, recall, and trust.
A strong oral paper does not stop at the formula. It explains:
Do not pile all numbers into one page or paragraph. Break results into:
Strong oral papers do not treat analysis as number recitation.
Use analysis to answer:
The prose before and after a figure or table should tell the reader:
In strong oral papers, main-text prose does not waste its budget by restating numbers the reader can already read from a table or plot.
Use displays for:
Use prose for:
When the display is a benchmark block, the prose may summarize the headline pattern, but it should not be the only place where the comparison surface exists.
If the evidence supports "strong default," "wins or ties most settings," or "more robust under sweep," do not escalate the wording into universal dominance.
Overclaiming wastes reviewer trust that the rest of the paper worked hard to build.
If you removed competitor rows, compressed the metric spread, or moved key comparison context out of view, narrow the comparative wording accordingly.
A method section can be principled and still overconsume main-text budget.
Compress repeated defense if that space is more valuable as:
An oral paper is usually defended by main text plus appendix together. Treat the appendix as part of the persuasion system, not as detached storage.
These are strong signals that a draft still reads like a compressed or LLM-like paper:
When using this skill, leave behind one or more of the following:
Prefer concrete edits over generic advice.
references/oral_package_patterns.mdreferences/oral_writing_principles.mdreferences/section_rewrite_checklist.mdreferences/experiments_analysis_patterns.mdartifact.arxiv(paper_id=..., full_text=False)paper-plot first. Use figure-polish only after a durable first-pass render exists. Sync resulting figure paths and takeaways back into paper/evidence_ledger.json, paper/paper_experiment_matrix.md, and the draft.nature-* skill only after the current section job, evidence rows, and unresolved fields are known. Use the companion skill to produce a bounded section/figure/deck deliverable, then return to write to integrate it into the draft, evidence ledger, figure/table catalog, references, and bundle status.Draft To Top Conference Oral section below.artifact.validate_manuscript_language(detail='full'), and artifact.validate_manuscript_coverage(detail='full'). A short memo is only artifact.submit_paper_bundle(package_type='draft_checkpoint', ...); use submission_package only when submission_ready=true.bash_execmemory.list_recent(...) and memory.search(...):
use at the start of substantial writing passes, before route changes, and before repeating search or drafting patterns that may already have reusable lessons.memory.write(...):
use only for reusable lessons such as citation retrieval rules, packaging traps, figure-integration lessons, or section-rewrite heuristics; do not store one-off draft text, transient wording, or current-section notes that should live in files.remaining work on this manuscript64 + 64 or 64+64 in manuscript prose, titles, abstracts, captions, or conclusions.paper/paper_experiment_matrix.jsonpaper/evidence_ledger.jsonpaper/evidence_ledger.mdresult_table rows, active evidence, or paper matrix rows disagree, stop drafting and repair the paper contract first.references/outline-evidence-contract-example.md and references/paper-experiment-matrix-template.md when rebuilding the contract. Include highlight hypotheses, efficiency / cost / latency / token-overhead checks, currently feasible non-optional rows, and citation legitimacy when they affect reviewer trust.paper/reviewer_first_pass.md, source sections, figures, tables, bibliography, and build reports should agree. Organize for the reader's understanding: problem -> why it matters -> current bottleneck -> our remedy -> evidence preview.https://github.com/ResearAI/AutoFigure-Edit, or https://deepscientist.Analysis should not continue the result dump. It should explain mechanism, trend, tradeoff, or failure behavior that the reviewer cannot infer from the visible numbers alone, and it should use a visible display or table when the interpretive claim depends on evidence the reader would otherwise not see.Analysis should remain a standalone reviewer-facing layer after headline results. Keep at least two visible check blocks, subsections, or strongly signposted units when the staged package separates mechanism, credibility, robustness, tradeoff, sensitivity, or failure-boundary work instead of collapsing everything into one short afterword.Analysis should own the headline validation burden when the paper first needs to prove that a metric, proxy, or diagnostic is meaningful before moving to interventions, recommendations, or downstream design guidance. Do not let analysis devolve into a leftover mechanism note if it is carrying primary credibility work in the staged evidence package.Analysis should keep a minimum main-text evidence floor before deferring support to the appendix: preserve at least one mechanism or credibility display and at least one tradeoff, robustness, sensitivity, or quality-support display when the staged package uses them to answer different reviewer concerns.Analysis should open with an explicit taxonomy, mechanism frame, or tradeoff frame when later interpretation depends on named categories. If the gold package distinguishes failure types such as programming, planning, and summarization, define those categories before interpreting shifts between them.Appendix should be written before limitations, conclusion, and abstract so later sections can accurately describe the support package that actually exists.Integration should check cross-section consistency, display roles, appendix bridges, and claim calibration, not rewrite the paper from scratch.Integration should remove meta-signposting or planning language that still reads like drafting scaffolding, and it should preserve one memorable qualitative, human, or failure anchor when the staged package can support it.Integration should check titles, abstract, captions, conclusion, and section openings for user/operator/route wording; these locations must read like paper text, not process notes.Integration should replace generic appendix mentions with precise labeled destinations whenever the body section already knows the supporting overflow lane.Integration should audit canonical section jobs, not just headings.