Imported from AIScientists-Dev/academic-humanizer (SKILL.md). Install upstream with npx skills add AIScientists-Dev/academic-humanizer. Copyright stays with the author (MIT).

Academic Humanizer

Improve the clarity and voice of AI-assisted academic writing while keeping the precise, evidence-bound voice that scholarship requires and matching the author's own style. It preserves every number, result, and citation, and it is not a tool for evading AI-use disclosure.

When to use

Editing or reviewing academic prose: paper sections, abstracts, rebuttals, related work, and funding proposals (NSF Project Summary/Description, NIH Specific Aims, fellowship and foundation proposals; see Layer 6). Not for blogs, marketing, or personal essays, and never inject opinion, humor, or first-person "personality" into a manuscript. For technical writing, neutral and precise is the human voice. One caveat for proposals: their register is different from a paper's, since they are sold on vision and feasibility, so the ambition language a paper would trim is appropriate there; apply Layer 6, not the paper layers' stricter trimming, to vision statements.

Core principle

Academic writing already has a correct human voice: neutral, precise, third-person plural ("we"), every claim tied to its evidence. The job is to (1) strip the AI tells without casualizing, and (2) enforce the discipline a general humanizer misses: every claim earns its number, figure, or citation, and no verb is stronger than its evidence.

Process

1. Read the manuscript and any author writing sample; note the document type (paper vs. funding proposal) and the target venue or funding agency. For proposals, also apply Layer 6 and preserve appropriate vision.
2. Audit (do not edit yet): list each detected pattern with its location and proposed fix, and each empirical claim's evidence status.
3. Rewrite: same structure and content, all claims and citations preserved, tells removed, over-claims matched to evidence, legitimate hedging kept.
4. Report: cleaned text plus a short change log (patterns removed, claims softened or given evidence pointers, voice notes). Cover everything the original covered: if it had five paragraphs, so does the rewrite.

Layer 1: General AI-tell catalog

Scan for and fix the general patterns, subject to the academic exceptions in Layer 3: inflated significance ("marking a pivotal moment"); superficial "-ing" tails that fake depth ("..., highlighting..."); promotional/figurative language ("rich", "vibrant", "groundbreaking"); vague attributions ("experts argue" with no cite); AI vocabulary (delve, underscore, intricate, tapestry, testament, landscape (abstract), pivotal, showcase, foster, leverage (filler), realm, seamless); copula avoidance ("serves as" -> "is"); negative parallelisms ("not just X, but Y"); rule-of-three padding; elegant variation (cycling synonyms for one referent); filler ("it is worth noting that", "in order to"); overlong, clause-stacked sentences (split them; see 2.11); and em-dashes (remove entirely; recast with commas, colons, parentheses, or separate sentences).

Before: Additionally, an enduring testament to the method's value is its ability to delve into intricate dependencies, showcasing a seamless integration that underscores its pivotal role. After: The method also captures higher-order dependencies, which the baselines miss (Table 2).

Layer 2: Academic AI tells (remove or fix)

2.1 Over-claiming verbs

Empirical work shows and provides evidence; it does not prove or demonstrate universal truths. Watch: demonstrate, prove, establish, confirm, guarantee; "significantly" with no test/number. Before: We prove that our method significantly outperforms all prior approaches. After: Our method improves held-out accuracy by 4--7 points over the strongest prior approach (Table 3); the gain is significant at p < 0.01 by a paired test.

2.2 Significance hype

Watch: paves the way for, a crucial/pivotal step toward, has the potential to revolutionize, opens new avenues, sheds light on, of paramount importance, bridges the gap. Before: This work paves the way for a new paradigm and sheds light on a problem of paramount importance. After: This work addresses one failure mode of prior methods: error accumulation under long-horizon rollout (Section 4).

2.3 Empty intensifiers

Watch: extensive/comprehensive/thorough experiments, a wide range of, numerous, various. Before: We conduct extensive experiments on a wide range of datasets. After: We evaluate on three datasets (ImageNet, CIFAR-100, and iNaturalist).

2.4 Novelty padding

Watch: "novel" used more than once per section; "to the best of our knowledge"; "for the first time". Before: We propose a novel framework and, to the best of our knowledge, are the first to study this. After: We study online calibration under delayed labels, which prior calibration work (offline) does not address.

2.5 Formulaic openers

Watch: "In recent years, X has attracted increasing attention"; "With the rapid development of..."; "Despite recent advances,...". Before: In recent years, tabular deep learning has attracted increasing attention. After: Tabular deep learning has a structural limitation: most models discard feature-type metadata and must relearn it from data.

2.6 Connective overuse

Do not start consecutive sentences with Moreover/Furthermore/Additionally/In particular; let logic carry. Before: Moreover, the method is fast. Furthermore, it is simple. Additionally, it scales. After: The method is fast and simple, and it scales to one million rows (Section 5).

2.7 Contribution-list cliches

Each contribution names a specific result, not a restatement of the abstract. Before: Our contributions are: (1) a novel method; (2) extensive experiments; (3) strong results. After: We (1) introduce a metadata-aware encoder that reaches 0.91 AUROC vs 0.86 for the strongest baseline; (2) show it stays within 2 points under 20% label noise where the baseline drops 9; (3) release the benchmark.

2.8 Citation dumping

Cite the one or two works that matter and say why, not a bracketed list. Before: Many methods exist [3, 7, 9, 12, 15]. After: The closest prior method is TabNet [7], which encodes all features jointly; we instead condition on feature-type metadata.

2.9 Hedging-by-vagueness

Watch: somewhat, relatively, fairly, to some extent, quite. Quantify or cut. Before: Performance is somewhat better and relatively robust. After: Accuracy is 3 points higher and varies by less than 1 point across five seeds.

2.10 Boilerplate emphasis

Watch: "It is worth noting that", "It should be emphasized that", "Notably,", "Importantly,". If it matters, the sentence shows it. Before: It is worth noting that, importantly, the gain holds across scenarios. After: The gain holds across all three scenarios (Table 4).

2.11 Overlong, clause-stacked sentences

AI favors long sentences that chain three or four clauses with commas and "which", "that", "while", "with". Split them: one idea per sentence, and cut subordinate clauses that carry no weight. Watch: sentences past ~30 words, or with 3+ subordinate clauses. Before: Existing methods, though promising, are largely empirical, with unclear principles underpinning their behavior, which limits their reliability and further progress. After: Existing methods stay empirical. Their principles are unclear, which limits reliability and progress.

Layer 3: Preserve these (do NOT over-correct)

A general humanizer flattens legitimate scholarly constructs. Keep them.

• Evidence-tied hedging is correct and required. Keep "suggests", "is consistent with", "we hypothesize that", "may indicate", "appears to" when the claim is genuinely uncertain. Wrong fix: turning "the results suggest X" into "the results prove X": this manufactures over-claiming. Keep the calibrated verb.
• Passive voice is fine when the actor is irrelevant: "Samples were normalized to total protein."
• First-person plural "we" is standard; do not rewrite to avoid it.
• Semicolons and an occasional triple are fine in moderation. Em-dashes are the exception: remove them entirely (Layer 1), recasting with commas, colons, parentheses, or separate sentences.
• Formal definitions, named methods/metrics, technical terms, equations, and symbols stay verbatim.
• Never invent, drop, or alter a number, equation, or citation. Same content; preserve every cite key.

Layer 4: Claim-evidence discipline

For every empirical claim, check (a) is it backed by a number, figure, table, or citation in the text, and (b) does the verb match the strength of that evidence?

• Unbacked claim -> add the evidence pointer or soften. Before: Our method is more robust. After: Our method's accuracy drops by 2 points under distribution shift, versus 11 points for the baseline (Figure 3).
• Verb stronger than evidence -> downgrade. Before: This demonstrates that our method is universally superior. After: On these three datasets, our method matches or exceeds the strongest baseline (Table 2).
• Vague magnitude -> a number or RANGE, attributed. Before: a large improvement. After: a 2--6% improvement in balanced accuracy over the strongest baseline. Prefer ranges (e.g., "2--6%") over single averaged values unless the averaging method is stated, and attribute each number to its method, metric, and baseline. When comparing, lead with the comparison against the strongest competitor, not the trivial baseline.

Layer 5: Voice and venue matching

If the author supplies prior papers, read a sample first and note sentence rhythm, connective habits, level and placement of hedging, how they open sections, notation, and recurring phrasings, then match them. Match the venue's register too (e.g., ICLR/NeurIPS: terse, direct, results-forward; Nature/PNAS: more expository). Absent a sample, default to clean, precise, venue-appropriate prose, not the casual, opinionated voice of a general-purpose humanizer.

Layer 6: Funding-proposal mode (NSF, NIH)

A proposal is not a paper. It is sold on vision plus feasibility, not on finished results, and reviewers score it. The register shift matters: ambition language that the paper layers would trim ("long-term goal", "pioneer", "transformative", "establish a foundation") is appropriate and expected here, provided a credible plan and evidence back it. So in proposal mode, do not flatten the vision; enforce a different discipline instead: claim <-> feasibility.

6.1 Know the structure; the score lives in the first pages

Reviewers form a score from the opening, then skim the rest to confirm it. Put most editing effort there.

• NSF. A one-page Project Summary with the three review-criteria heads spelled out: Overview, Intellectual Merit, Broader Impacts, each self-contained. The Project Description then opens with long-term vision -> this proposal's goal -> the gap -> the specific thrusts/aims -> the payoff, ideally within the first 1--2 pages, with one overview figure. Broader Impacts must be substantive and integrated, never an afterthought.
• NIH (R01). The Specific Aims page is the whole proposal in one page, and is the most-read, most-decisive page. Standard arc: (1) opening: the problem, what is known, the gap / critical need; (2) the long-term goal and the central hypothesis with its rationale; (3) "The objective of this application is..." plus how the hypothesis was formed; (4) 2--3 Aims, each a one-line goal + a phrase on approach + the expected outcome; (5) a payoff paragraph: what changes if it succeeds. Then Significance, Innovation, Approach as separately scored sections.

6.2 First-3-pages primacy (edit these hardest)

By the end of page 1 (NIH Aims) or pages ~2--3 (NSF), the reader must already hold: the hook (why it matters, concretely), the gap (what is missing and the cost of the gap), the central idea (your approach in one sentence), the aims/thrusts (crisp and parallel), and the payoff. If any is missing or buried, fix that before touching later sections. A reviewer unconvinced by page 3 does not recover on page 10.

6.3 Proposal-specific weak moves to fix

• Vague importance. Watch: "this is an important/timely problem", "X has many applications". Before: Understanding this problem is critically important. After: Without bounds on how measurement noise propagates to diagnosis, clinical models are tuned by trial and error, the inefficiency this proposal removes.
• Method-as-aim (an aim naming a technique instead of a question or outcome). Before: Aim 2: Apply transfer learning to the dataset. After: Aim 2: Determine whether fusing wearable and lab signals improves early detection, and for which patient subgroups it helps or hurts.
• Dominoed aims (Aim 2/3 collapse if Aim 1 fails; reviewers flag this as fragile). Fix: phrase aims as parallel and independently valuable; where one depends on another, state the fallback.
• Ambition without feasibility. Every bold claim needs a footing: preliminary data, a prior result/publication, a classical theorem you build on, or a collaborator/letter. Fix: attach the evidence beside the claim ("our preliminary result in Fig. X shows...", "building on a classical minimax lower bound...").
• Boilerplate Broader Impacts / training plan. Watch: "we will mentor students and disseminate via talks and papers." Fix: make it concrete, enumerated, and tied to the research: specific programs, named courses or tools, measurable outreach.
• Hedged central hypothesis. The Aims-page hypothesis is a falsifiable commitment, not "we will explore whether possibly...". Calibrated hedging belongs in the Approach's interpretation, not the central claim.

6.4 Preserve and deploy (funded-proposal craft)

These read as strength; keep or add them rather than editing them out.

• Vision/ambition framing: a bold long-term goal up front, with this proposal as one principled step toward it.
• Run-in lead-ins for scannability: bold/italic Goal:, Motivation:, Innovation:, Thrust/Aim N (one-line mission):. Reviewers skim; visible structure earns time.
• A concrete running example or protagonist to make an abstract method vivid and consistent across aims.
• Sharp challenge/aim statements posed as questions: a crisp open question reads as a well-posed problem (a boxed or set-off question per aim works well).
• Anchoring novel work in deep, named classical results to signal rigor and lineage: a known inequality, capacity notion, or test that the new method generalizes.
• Foreground the team's standing as feasibility evidence: prior funded work, preliminary results, publications, collaborators, and demonstration partners belong early, as proof the plan is executable. A real track record is evidence, not boasting; place it where it de-risks the aims. (Use only the PI's own real, supplied record; never invent funding, results, partners, or letters.)

6.5 Claim <-> feasibility (the proposal analog of Layer 4)

For every aim and promised outcome, check: is the ambition matched by a credible means, such as preliminary data, a prior method, a classical foundation, a collaborator, or staged de-risking? If yes, keep the ambitious verb. If no, attach the missing evidence or scale the claim to what the plan supports. Never invent preliminary results, prior funding, partners, or letters; if the support does not exist, flag the gap for the author rather than papering over it.

Output

Return the cleaned text plus a short change report: patterns removed (by type), claims softened or given evidence pointers, and any voice/venue notes. Confirm that no number, equation, or citation was altered.