科研 Skill

计算机论文写作助手

computer-science-writer

根据已有文献综述与实验结果撰写可审计计算机论文

官方收录发布者 local安全审核通过v0.1.0更新于 2026-07-15

Skill 说明

基于已有文献综述、方法说明和实验结果撰写可审计的计算机论文,建立主张—证据映射,组织计算机论文结构、表图与结果讨论,生成 LaTeX 权威稿并通过 tex_to_docx 导出 DOCX。 | 输入:文献综述、references.bib、方法/系统说明、实验表格/日志/图片、可选投稿模板 | 输出:source_inventory.md、writing_constraints.json、claim_evidence_map.json、paper_outline.md、mainbody.tex、paper_audit.json、manuscript.docx

SKILL.md

Computer Science Writer

Turn supplied literature and completed experiment artifacts into a defensible paper. Treat the files, not a plausible narrative, as the authority.

Use this spine:

literature review + experimental artifacts
-> source inventory
-> claim-evidence map
-> paper architecture
-> tables/figures and Results
-> full manuscript
-> audit
-> mainbody.tex
-> tex_to_docx DOCX export

Boundaries

  • Start after the user has a topic or system/method and experimental results. Do not originate a new research idea.
  • Do not rerun, tune, replace, or extend experiments unless the user explicitly asks. Record missing evidence as a limitation or TODO.
  • Do not invent metrics, baselines, datasets, sample sizes, hardware, hyperparameters, p-values, effect sizes, citations, quotations, or contribution claims.
  • Infer organizing research questions only from the stated goal and supplied artifacts. Do not use an inferred question to broaden the study.
  • Distinguish an observed result from an explanation. Label explanations as interpretations unless the supplied experiment directly tests them.
  • Use calibrated verbs. Prefer shows, is associated with, or is consistent with; use causes, proves, outperforms, state of the art, or first only when the evidence supports the exact scope.

Conditional References

Read only the references needed for the active stage:

  • Read references/input-and-evidence.md when inventorying inputs or building claim_evidence_map.json.
  • Read references/paper-architecture.md before outlining, drafting, revising, or reviewing prose.
  • Read references/results-and-discussion.md before writing from quantitative or qualitative results.
  • Read references/tables-and-figures.md before creating, reformatting, captioning, or citing tables and figures.
  • Read references/citations-and-formatting.md when handling literature, BibTeX, venue style, LaTeX, or cross-references.
  • Read references/latex-and-docx.md before producing the canonical LaTeX manuscript or Word deliverable.
  • Read references/output-audit.md before final delivery and run both bundled validators.

Required Workflow

Follow all stages for a full-paper request. For a local request, such as revising only the Results section, execute the applicable stages and preserve the same evidence rules.

Stage 1. Confirm Writing Inputs

Locate and inspect:

  • the paper goal, method/system description, or existing outline;
  • the literature review and its citation records (.bib, DOI list, notes, or cited papers);
  • experiment outputs such as CSV/JSON logs, tables, plots, statistical reports, human-study materials, or an existing results draft;
  • optional venue template, author guidelines, page limit, language, and anonymity requirements.

For a full paper, require both literature material and experiment evidence. If either is absent, stop full-paper drafting and ask for it. Continue with a clearly scoped partial task when the user requests one.

Never treat a screenshot or prose summary as more precise than an available machine-readable result. When sources conflict, report the conflict before drafting the affected claim.

Create source_inventory.md with each input's path, type, role, scope, and known limitations. Do not copy or modify raw result files.

Stage 2. Record Constraints

Create writing_constraints.json:

{
  "language": "en",
  "target_venue": null,
  "article_type": "conference|journal|thesis|technical-report",
  "page_or_word_limit": null,
  "citation_style": "venue-template|IEEE|ACM|author-year|unspecified",
  "anonymized": false,
  "latex_template": null,
  "docx_required": true
}

Apply a supplied venue/template as the authority. If no venue is specified, use a conservative generic CS article structure and semantic LaTeX; record assumptions instead of blocking progress.

Stage 3. Build The Claim-Evidence Map

Convert the existing study goal and results into research questions and bounded claims. Read references/input-and-evidence.md first.

Create claim_evidence_map.json. Give every claim:

  • a stable claim ID;
  • exact source file paths and locators such as row, column, table, figure, sheet, or log key;
  • reported statistics and comparison scope;
  • caveats and permitted wording;
  • intended section and table/figure IDs.

Validate it before drafting:

python3 /home/ubuntu/speedclaw_skillsurvey/ComputerScience_writer/scripts/validate_claim_evidence_map.py \
  claim_evidence_map.json --project-root .

Reject or downgrade claims without traceable evidence. Literature can support context and interpretation; it cannot substitute for missing experimental evidence.

Stage 4. Design The Paper Argument

Read references/paper-architecture.md. Produce paper_outline.md that maps every section to:

  • its reader question and function;
  • the claim IDs it may express;
  • the citations, tables, and figures it uses;
  • its expected length under the venue limit.

Derive the contribution list from demonstrated deltas: a new method/system artifact, measured performance or efficiency, empirical insight, dataset/benchmark, design implication, or reproducibility asset. Remove contributions that lack evidence in the map.

Stage 5. Normalize Tables And Figures

Read references/tables-and-figures.md. Reuse supplied quantitative outputs and regenerate them only from supplied data or scripts when necessary for consistent formatting.

  • Use stable labels such as tab:main_results, fig:architecture, and fig:latency_cdf.
  • Make every comparison disclose metric direction, units, dataset/task, aggregation, variability, and number of runs or participants when available.
  • Do not convert absent uncertainty into fake error bars.
  • Keep quantitative graphics programmatic; do not use image-generation models for empirical axes or values.
  • Explain missing baselines or incomplete cells in notes rather than hiding them.

Stage 6. Draft In Evidence-First Order

Draft in this order unless the user requests otherwise:

  1. Evaluation/Experiments and Results from verified outputs.
  2. Method/System/Approach from supplied design artifacts.
  3. Discussion, limitations, threats to validity, ethics, and reproducibility.
  4. Related Work from the supplied review and citation library.
  5. Introduction and contributions after the evidence boundary is stable.
  6. Abstract and title last.

Read references/results-and-discussion.md for Results paragraphs. Keep method, result, and interpretation distinct. Attach every numeric or comparative statement to a claim ID during drafting; claim IDs may live in comments or the audit manifest and must not appear in the polished prose.

Stage 7. Build The Canonical LaTeX Manuscript

Read references/citations-and-formatting.md and references/latex-and-docx.md.

Generate:

manuscript/
  mainbody.tex
  references.bib
  supplementary.tex
  figures/
  manuscript.docx          # when requested and export succeeds
  docx_export/             # export configuration and template artifacts

Use manuscript/mainbody.tex as the source of truth for prose, equations, citations, labels, tables, captions, and legends. Preserve a supplied venue class and template. Use BibTeX keys from the supplied or verified bibliography; never create a citation solely because a sentence “needs one.”

Stage 8. Audit Before Delivery

Read references/output-audit.md. Run:

python3 /home/ubuntu/speedclaw_skillsurvey/ComputerScience_writer/scripts/validate_claim_evidence_map.py \
  claim_evidence_map.json --project-root .

python3 /home/ubuntu/speedclaw_skillsurvey/ComputerScience_writer/scripts/audit_manuscript.py \
  --tex manuscript/mainbody.tex \
  --bib manuscript/references.bib \
  --evidence-map claim_evidence_map.json \
  --project-root . \
  --report manuscript/paper_audit.json

Treat validator errors as delivery blockers. Fix the manuscript or explicitly document an evidence gap; do not weaken the validator or silently remove failed claims.

Stage 9. Export DOCX Through tex_to_docx

Use only the existing local skill at:

/home/ubuntu/speedclaw_skillsurvey/tex_to_docx

For journal-aware export:

python3 /home/ubuntu/speedclaw_skillsurvey/tex_to_docx/scripts/tex_to_docx.py all \
  --source manuscript/mainbody.tex \
  --output manuscript/manuscript.docx \
  --guidelines-url "<target venue author guidelines URL>" \
  --journal "<Target Venue>" \
  --language en \
  --work-dir manuscript/docx_export \
  --bib manuscript/references.bib \
  --toc

If a venue configuration and reference DOCX already exist, use the staged export command described in references/latex-and-docx.md. Check for Pandoc first. If unavailable, deliver the audited LaTeX sources and report that DOCX conversion is blocked; do not replace the exporter with an ad hoc Word generator.

Deliverables

For a full paper, deliver:

  • source_inventory.md
  • writing_constraints.json
  • claim_evidence_map.json
  • paper_outline.md
  • manuscript/mainbody.tex
  • manuscript/references.bib
  • manuscript/supplementary.tex when applicable
  • copied or reproducibly generated manuscript figures
  • manuscript/paper_audit.json
  • manuscript/manuscript.docx when requested and export succeeds

For a partial writing request, deliver the requested section plus the relevant evidence map and audit results.

Non-Negotiable Quality Gates

  1. Trace every number and comparison to a supplied result artifact.
  2. Resolve every citation key to real bibliography metadata.
  3. State the correct comparison scope; do not generalize one dataset, workload, language, population, or device to all settings.
  4. Report negative, null, and mixed results honestly.
  5. Separate performance, efficiency, robustness, usability, privacy, fairness, and causal claims; evidence for one does not establish another.
  6. Match every table/figure reference to an existing labeled object and every displayed object to discussion in the text.
  7. State run counts, uncertainty, statistical tests, and multiple-comparison handling when supplied; explicitly mark them unavailable when they were not reported.
  8. Include limitations and threats to validity proportional to the design.
  9. Keep anonymous submissions free of author names, affiliations, acknowledgments, repository identities, and self-identifying prose when required.
  10. Deliver Word only from the audited canonical LaTeX source through tex_to_docx.

常见问题

计算机论文写作助手能做什么?

基于已有文献综述、方法说明和实验结果撰写可审计的计算机论文,建立主张—证据映射,组织计算机论文结构、表图与结果讨论,生成 LaTeX 权威稿并通过 tex_to_docx 导出 DOCX。 | 输入:文献综述、references.bib、方法/系统说明、实验表格/日志/图片、可选投稿模板 | 输出:source_inventory.md、writing_constraints.json、claim_evidence_map.json、paper_outline.md、mainbody.tex、paper_audit.json、manuscript.docx。

如何安装计算机论文写作助手?

在科搭 Codar 左侧点击「插件」,在「组件」页签搜索「计算机论文写作助手」并安装;在 Claude Code 或 Codex 中可以发送本页的安装提示词,或下载 Skill 包后解压到 ~/.claude/skills 或 ~/.codex/skills。

计算机论文写作助手需要哪些权限?

该科研 Skill声明需要:本地文件。安装前会再次列出,由你确认后才会生效。

计算机论文写作助手经过官方审核吗?

是。SpeedSkills 收录的每个版本都会记录内容摘要与权限声明,当前版本已通过安全审核。