human-academic-writing

name

human-academic-writing

description

Revise English academic manuscript prose so it reads like careful human scholarship rather than generic AI-polished text. Use when Codex drafts, translates, rewrites, edits, polishes, compresses, or de-AI's research-paper text, LaTeX passages, abstracts, introductions, related work, methods, results, discussion, conclusions, captions, reviewer responses, or any manuscript section where the user asks for a less AI-sounding, more natural, precise, evidence-bound academic voice. Use alongside research-paper-writing, academic-research, and academic-manuscript-agents when those broader skills also apply.

Human Academic Writing

Overview

Use this skill as a style and revision layer for English academic writing. Its job is not to make text ornate or detector-proof; its job is to make prose specific, claim-led, evidence-bound, and recognizably authored by a careful researcher.

Core Workflow

1. Identify the passage role: abstract, introduction, related work, method, results, discussion, conclusion, caption, response letter, or local paragraph polish.
2. Read the surrounding context when available: section goal, target venue, figure/table/data claims, citations, and project instructions.
3. Diagnose before rewriting. Mark the main issue as one or more of: inflated claim, vague noun, filler transition, translationese, robotic symmetry, unsupported evidence, weak paragraph role, or venue mismatch.
4. Rewrite at the smallest useful scope. Preserve the user's meaning, LaTeX, citations, equations, labels, terminology, and measured results.
5. Run a style and evidence audit before returning prose. If a claim lacks support, weaken it or flag the missing evidence instead of smoothing over the gap.

Reference Routing

Load only the reference needed for the current task:

• references/anti-ai-style.md: use for "too AI", "humanize", "de-AI", "natural academic English", Chinese-to-English translation polish, sentence-level style, and paragraph-level polish.
• references/section-craft.md: use for drafting or revising full sections, especially abstracts, introductions, related work, methods/design, experiments/results, discussion, conclusions, captions, and headings.
• references/revision-gates.md: use for full-passage checks, final polish, pre-submission review, or any task where a patch must be justified by diagnostics.

Non-Negotiables

1. Evidence beats elegance. Do not add claims, citations, numbers, datasets, baselines, implications, or venue facts not present in the user's material or verified sources.
2. Do not churn natural text. If the passage already sounds human and precise, say so and only fix concrete issues.
3. Prefer named mechanisms, metrics, datasets, baselines, and conditions over generic adjectives.
4. Keep claim strength calibrated: what the evidence shows, what it suggests, where it applies, and where it does not.
5. Preserve technical meaning and markup. Keep LaTeX commands, math, citations, references, labels, and formatting unless the user explicitly requests structural editing.
6. Follow any project safety rules before editing files. For existing manuscript files, propose patches and wait for explicit approval when required.

Default Voice

• Plain but not flat.
• Specific but not overloaded.
• Confident when evidence is direct, cautious when evidence is partial.
• Claim-first at paragraph openings when the section allows it.
• Active voice when agency or method matters; passive voice only when convention, focus, or unknown agency justifies it.
• Varied sentence length, but no decorative complexity.
• Transitions should explain relationships, not decorate the paragraph.

Task Modes

For de-AI or humanizing:

1. Load references/anti-ai-style.md.
2. Identify exact AI-like symptoms.
3. Rewrite only the affected sentences or paragraphs.
4. Return the revised passage plus a short note on the main fixes.

For section drafting or major revision:

1. Load references/section-craft.md.
2. Build paragraph roles or topic sentences before full prose.
3. Map each major claim to evidence or a needed source.
4. Draft in paragraphs, then run references/revision-gates.md.

For translation into English academic prose:

1. Preserve the author's technical intent before improving idiom.
2. Avoid literal Chinese noun stacks and empty evaluative adjectives.
3. Keep terms, symbols, citations, and result values stable.
4. If a concept has no clear English equivalent, offer one preferred rendering and one alternative.

For reviewer responses:

1. Separate manuscript wording from response-letter wording.
2. Use direct, respectful prose.
3. State the change and where it appears.
4. Do not imply the reviewer misunderstood unless the revised manuscript now removes the ambiguity.

Output Contract

For chat-only edits, return:

1. Diagnosis: one to four concrete issues.
2. Revised text: clean manuscript-ready prose.
3. Notes: only source/evidence gaps, claim-strength changes, or terminology choices that matter.

For existing-file edits, return:

1. Affected file list.
2. Short rationale.
3. Proposed patch or explicit edit list.
4. Evidence or citation gaps.
5. Wait for explicit user approval when project rules require it.