Imported from lengmo1996/VibeResearch-skills (skills/global/writing-academic/SKILL.md). Install upstream with npx skills add lengmo1996/VibeResearch-skills --skill writing-academic. Copyright stays with the author.

Academic Writing

Purpose

Create or revise academic manuscript prose without changing scientific content. Use literature-synthesis when finding or comparing literature is primary, writing-manuscript-audit when the requested deliverable is a reverse outline, defect list, or verdict, and research-result-analysis when interpreting raw results is primary.

Required and optional inputs

Required: writing goal, output language, and either target text or verified content points. Optional: venue/style, length, audience, manuscript context, terminology/symbol table, citations, evidence map, reviewer comments, constraints, and an author-owned voice sample or approved task-scoped voice profile. A de-template handoff may also include accepted AUD-* findings, protected content, and verification methods. revision-patch additionally requires a contained target manuscript, stable block markers, exact pre-edit and reviewed replacement hashes, replacement artifacts, and explicit authorization before application.

Workflow

1. Bind the source material, requested operation, output language, and immutable items: claims, numbers, formulas, metric direction, citation keys, quotations, URLs, code, and user-designated text.
2. Select the smallest sufficient operation. Do not run a full rewrite for a grammar, translation, or local-clarity request.
3. For an explicit whole-paper outline or structural draft, load full-manuscript lifecycle, use paper-outline, and create or patch the shared paper-context only when a cross-section state index is useful. For a scoped abstract, introduction, or method outline or structural draft, instead load exactly one matching guide: abstract, introduction, or method, then use section-structure-plan.md. Do not load the full-paper and section guides together, and do not load either for local rewrite, grammar, translation, or naturalness-only requests.
4. Build or reuse terminology and symbol decisions. Identify claim-to-evidence gaps before drafting and mark unsupported claims [citation needed].
5. Write at the requested structural level and claim strength. For related work, require a literature-synthesis evidence map and bind each comparative claim.
6. For de-template, read naturalness rewrite, bind accepted AUD-* findings when supplied, and run its one-pass rewrite and residual check. For author-voice, instead read author voice calibration and use only the user-owned task-local sample or approved profile. Do not load both references unless both outputs are explicitly requested; do not use either for authorship detection or detector optimization.
7. For compress, bind a measurable target when one is supplied, remove redundancy before evidence, and verify the before/after length plus protected scientific content. Use compression-report.md when the user requests an audit trail or the source is high risk.
8. Verify that meaning, data, formulas, references, uncertainty, protected spans, and source attribution are preserved. For high-risk text or an auditable handoff, use protected-spans.json and run <python-command> <skill-root>/scripts/verify_protected_spans.py --source <before> --revised <after> --manifest <protected-spans.json>.
9. For revision-patch, read revision patch protocol, create revision-patch.json, and run <python-command> <skill-root>/scripts/apply_revision_patch.py --root <workspace-root> --manifest <revision-patch.json>. This default validates and previews only. Add --apply solely when the user authorized that exact target; any target, block, path, or hash mismatch must stop without writing.
10. For outline, return the story map and paragraph plan. For prose operations, return revised text first. Follow with only useful decisions and unresolved evidence.

Modes

• outline, draft, rewrite, de-template, author-voice, compress, expand, clarify, venue-style, reviewer-response, revision-patch.
• Section targets may include abstract, introduction, method, experiments, discussion, limitations, conclusion, and related-work.

RAG policy

Ordinary draft/rewrite is optional; related work is required; terminology consistency is recommended; reviewer response is optional; pure translation, grammar, formatting, or user-restricted-source editing is never.

Evidence policy

Never modify data, formulas, metric direction, citation keys, or supplied factual claims. Do not invent experiments, references, names, dates, numbers, or quotations. Candidate citations remain candidates until verified. Do not copy or closely imitate long source passages or imitate a named author's distinctive style.

Output contract

For a whole-paper outline, output the paper story, section dependency order, stable claim IDs, evidence gaps, and locked/open decisions; emit only a task-local paper-context patch unless persistence is separately authorized. For a section outline, output the story map and paragraph plan. For prose operations, output the revised text first. Add only useful sections: Change rationale; Terminology decisions; Claims requiring evidence; Candidate evidence; Unresolved questions. For related work, include or reference the evidence map. For compress, report target attainment and preservation failures when a compression report is required. For a naturalness rewrite, use rewrite-invariant-report.md only when an audit trail is useful; otherwise keep the checks internal. For de-template, preserve incoming AUD-* IDs and state which accepted findings were resolved, unchanged, or blocked. For author-voice, state the sample/profile scope and confidence without persisting it. A protected-span validator pass proves literal preservation only; it never certifies semantic equivalence. For revision-patch, return the manifest, target/source hash, per-block validation, protected-span result, preview or apply status, output hash, and unresolved conflicts. Validation success is not write authorization. An applied patch must preserve all bytes outside declared block interiors and use atomic replacement.

Failure behavior

If evidence required for a claim is unavailable, keep a neutral supported claim or mark [citation needed]; never fill the gap from memory. If the input is too incomplete for a requested draft, provide a scaffold and list missing verified facts. If a voice sample is too short or inconsistent, state that voice matching is low-confidence and use the requested academic register. If no material naturalness problem is present, preserve the text and report that no rewrite is needed. If a protected-span check fails, do not present the revision as final: restore the span or return the candidate with the failed IDs. Perform at most one corrective rewrite pass; do not loop against an AI detector or style score. If a revision target, marker, replacement, or hash differs from the manifest, stop with a conflict report. Do not rebase, fuzzy-match, overwrite, or silently regenerate hashes. Rebuild the manifest from the newly reviewed target instead.

Composition rules

May follow literature-synthesis or precede writing-manuscript-audit. Use no more than two supporting Skills. Reviewer response mode does not replace comment prioritization by writing-review-triage.

Examples

• “把这个 introduction 改成更清晰的学术英语，不改变引用和数据。”
• “根据这些已接受的 AUD findings 去掉模板化表达，返回修改后的段落。”
• “参考我提供的本人论文样本改写这段，保持术语和引用不变。”
• “根据这些已验证的贡献点，规划 introduction 的故事线和段落功能。”
• “根据已有 evidence map 写 related work。”
• “把已接受的 AUD-001 修订做成带原文 hash 的可审查 patch，先校验不要应用。”

Non-examples

• “核验这些引用是否支持 claim。” → writing-manuscript-audit.
• “反向梳理这篇论文的段落功能并找出结构缺陷。” → writing-manuscript-audit.
• “判断这段有多大概率由 AI 生成。” → unsupported authorship classification, not this Skill.
• “改写到可以绕过 AI detector。” → refuse detector-evasion optimization.
• “解释这些实验为何提升。” → research-result-analysis.

Validation checklist

• Meaning, data, formulas, citations, and uncertainty are preserved.
• Mode and audience are explicit.
• Structural references were loaded only for the matching section and operation.
• Whole-paper work loaded the lifecycle reference without preloading every section guide.
• A paper-context patch contains only thin state and artifact references and is not persisted implicitly.
• An outline has one story function per paragraph and explicit claim-evidence gaps.
• Unsupported claims are marked or weakened.
• Related work has an evidence map and required RAG.
• Terminology and symbols are consistent.
• No long-source imitation or fabricated content appears.
• Naturalness work preserved protected spans and passed a residual-pattern check.
• De-template work consumed accepted findings without rerunning a full audit.
• Author-voice work used only an author-owned task-local sample or approved profile and did not persist it.
• High-risk rewrites passed literal protected-span validation or exposed the failed span IDs.
• A low-signal naturalness request was allowed to return no change needed.
• Compression measured the requested target and preserved protected scientific content.
• Revision patches use contained paths, unique block IDs, exact target/block hashes, atomic apply, and fail closed on drift.
• Patch validation and patch application are reported separately; validation never implies write authorization.

Shared contracts and stop conditions

Follow the canonical RAG, evidence, failure, outputs, style, terminology, citations, paper context, claim evidence, platform compatibility, and file safety contracts. Stop when the requested text and checks are complete, required evidence is unavailable, or continuing would change unsupported scientific content or exceed the authorized artifact scope.