Academic Humanizer
Strips AI-writing tells from papers and NSF/NIH grant proposals while preserving scholarly voice and tying every claim to evidence.
SKILL.md declares a least-privilege toolset (Read/Write/Edit/Grep/Glob/AskUserQuestion), an audit-before-edit workflow with a change report, and README states clear ethical boundaries (no altering numbers/citations, not for evading AI disclosure); data flow is largely transparent. Deducted for: no explicit user confirmation gate before writes, no rollback mechanism described, and unverified publisher identity.
Instructions are self-consistent: six layers, audit→rewrite→report loop, preserve-rules do not contradict fix-rules, examples cover major patterns. Deducted for: no tests, thin handling of abnormal inputs (missing author sample, non-English text, malformed files), and weak failure feedback; static read caps score at 10.
Use and non-use boundaries are explicit (academic papers and NSF/NIH proposals, not blogs/marketing), trigger description is precise, and NSF/NIH structural knowledge is concrete. Deducted for: English-only scope with no Chinese-language support and no declared environment-fit notes, limiting fit for Chinese-language users.
Docs are well layered (frontmatter, layered body, README, examples, upstream attribution, MIT LICENSE), progressive disclosure is reasonable. Deducted for: version mismatch (SKILL.md 0.3.3 vs README badge 0.3.2), no changelog, unclear maintenance ownership/update path, and license metadata reported as NOASSERTION.
The core task (strip AI tells, preserve scholarly conventions, claim-evidence alignment) has a clear path and concrete before/after examples; marginal value over generic humanizers is well argued. Deducted for: no executed verification, output quality depends on model compliance, and the referenced examples file was not available in this review.
Auditable primary material exists: layered rules, before/after examples, upstream acknowledgments (blader/humanizer, ARMS), official NSF/NIH references, with claims vs. inference separated. Deducted for: no test suite, no CI, no independently reproducible results; evidence types are thin, and the static cap is 5.
- This is a static source-only review; nothing was executed and confidence is low.
- SKILL.md version 0.3.3 conflicts with the README badge (0.3.2) and there is no changelog; note the versioning gap.
- Repository license metadata is NOASSERTION while the LICENSE file is MIT; verify the license labeling on distribution channels.
- The skill rewrites manuscript files directly and does not require pre-edit confirmation or describe rollback; run on a copy first.
- The skill targets English academic writing only; no Chinese support. Core function has no overseas-service dependency and is usable offline from mainland-China networks.
- No tests or CI evidence; the guarantee that numbers and citations are preserved rests solely on instructions—verify manually after each use.
What does this skill do, and when should you use it?
Academic Humanizer is an editing-focused Agent Skill for AI-assisted academic prose (papers, theses, rebuttals) and funding proposals (NSF Project Summary/Description, NIH Specific Aims). It works in six layers: a general AI-tell catalog, academic-specific tells, protected scholarly conventions, claim-to-evidence matching, author voice and venue calibration, and a dedicated proposal mode. It never changes a number, result, or citation, and it is explicitly not designed to evade AI-use disclosure. It suits researchers who draft with AI but refuse to let their manuscripts read like everyone else's.
Reads the manuscript (paper sections, abstracts, rebuttals, or NSF/NIH proposal pages) and an optional writing sample of the author's prior work; audits first, listing each detected pattern with its location and each empirical claim's evidence status; then rewrites under the six-layer rules, checking that every claim is backed by a number, figure, or citation and that no verb outruns its evidence. Proposal mode additionally enforces claim-to-feasibility matching and first-page primacy. Output is the cleaned text plus a short change log (patterns removed, claims softened or given evidence pointers, voice notes).
- A PhD student or postdoc whose AI-drafted paper section needs 'paves the way' / 'delve' tells removed and over-claims matched to Table 3 before submission
- A faculty member preparing an NSF CAREER application who must polish a one-page Project Summary with Overview / Intellectual Merit / Broader Impacts headings
- An applicant writing an NIH R01 Specific Aims page who wants the central hypothesis checked for falsifiability and aims phrased as parallel, independently valuable
- A student polishing a rebuttal who wants polite hedging preserved while only generic AI boilerplate is stripped
- A PI standardizing a lab's writing style by feeding prior accepted papers as voice samples
What are this skill's strengths and limitations?
- Six layers cover the full editing pipeline from generic AI tells to NSF/NIH structure, with abundant real before/after examples including a funded NSF CAREER summary
- Explicitly preserves legitimate scholarly constructs (calibrated hedging, passive voice, 'we', definitions and symbols), avoiding the flattening that generic humanizers cause
- Strict never-touch policy on numbers, equations, and citations, plus claim-evidence matching on every sentence
- MIT-licensed ruleset designed to be forked and personalized to your own voice and field
- Voice matching depends on the model's judgment; there is no automated evaluation of fidelity to the author's style
- The GitHub license field reads NOASSERTION (README and SKILL.md state MIT); verify licensing yourself
- Rules reflect one research group's voice and may not fit every discipline or non-English writing
- Layer 6 distills stable NSF/NIH structure; current binding requirements (page limits, formatting) must still be checked against official sources
- No test suite; effectiveness is demonstrated by examples rather than benchmarks
How do you install this skill?
git clone https://github.com/AIScientists-Dev/academic-humanizer ~/.claude/skills/academic-humanizer — it is a plain SKILL.md plus examples, so Claude Code picks it up directly; Codex and MorphMind users can point their agent at SKILL.md as a skill or system prompt.
How do you use this skill?
In Claude Code, enter: /academic-humanizer, then paste a section or point at main.tex; optionally add "match my voice from prior_paper.pdf; target venue: ICLR". The skill runs an audit pass, then rewrites, then returns a change report.
How does this skill compare with similar options?
Compared with blader/humanizer (MIT, general AI-tell removal for blog and casual text), this skill reuses that Layer 1 catalog but extends it with academic evidence discipline and a proposal mode. Compared with koaeraser/ARMS (an autonomous idea-to-manuscript pipeline for statistics/methodology papers), this is the narrower, single-pass editing tool.