academic-humanizer is one prompt file, not a tool

The entire deliverable is one prompt file

The repository root holds .gitignore, LICENSE, README.md, SKILL.md, an assets directory and an examples directory, and the single example file named in the tree is examples/before-after.md. There is no package manifest, no build step and no source directory, which is why no primary language is detected for the project at all. What you install is prose that tells an agent how to edit. The install path is a clone straight into an agent's skills directory:

git clone https://github.com/AIScientists-Dev/academic-humanizer ~/.claude/skills/academic-humanizer

Because it is a plain SKILL.md plus examples rather than a compiled binary, the README says the same file also runs as a skill or a system prompt for Codex and MorphMind, and the instruction is simply to point your agent at SKILL.md. That portability is the design: the behavior lives in the prompt, and the hosting agent supplies the model, the session and the file access.

Calibration came from diffing drafts against accepted papers by hand

The rules were not invented. The team says it had the AI compare its own drafts against the team's accepted papers and funded proposals, then worked through the differences by hand, and it describes the result as nothing fancy. That provenance explains both the strengths and the limits of the prompt. Its instincts are tuned to how one group writes grant prose, which is why the README tells you to fork the repository and adapt the rules to your own: point it at a few of your past papers, keep the checks that fit your field, and adjust the rest. The stated goal is a personalized editing pass rather than a filter that fits everyone. It also explains the original motivation, since the complaint was never that AI drafts were wrong, but that they came out generic and verbose, opened with stock framing about recent years, drifted from the author's own voice, and lost the precision the work depends on.

The first disclaimer is about disclosure, not detection

The ethics section is short and worth reading in full before anything else. The README says the skill does not generate findings, invent data, or change citations, and that it is not designed to evade AI-use detection. It then states the part that matters most: using it does not remove your obligation to disclose AI assistance, and you should always follow the disclosure policy of the venue you submit to. The why-we-built section says the same thing in different words, ruling out gaming review, defeating detectors, and adding fake novelty as goals. What remains is an editing aid for clarity and voice, calibrated against an author's own prior accepted work. Framed that way, the interesting question is whether tightening prose helps a reviewer read the science, and everything else about the project follows from that being the only claim being made.

The first of six layers is credited to another MIT project

The prompt is organised as six layers: a general AI-tell catalog, then academic-specific tells, then preserving scholarly conventions, then matching claims to evidence, then voice and venue calibration, and finally a funding-proposal mode covering NSF and NIH structure, first-page primacy, and claim to feasibility. The acknowledgements are unusually specific about the first layer. The general catalog is reused from blader/humanizer, an MIT licensed project aimed at blog, casual and encyclopedic text, and extended here for academic prose. A second project, koaeraser/ARMS, an autonomous pipeline for statistics and methodology papers, informed the emphasis on claim to evidence matching and numerical precision. So one of the six layers is inherited rather than written here, and the README says so instead of leaving a reader to assume originality. The audit to rewrite loop that ties the layers together is defined in SKILL.md rather than in the README.

Grant mode concentrates on the pages reviewers actually score

Proposals get their own mode with different priorities from a paper. It keeps the visionary framing that a paper would trim, and it spends most of its effort on the first pages, on the reasoning that those are what reviewers score. The claim to feasibility check in that last layer is the mirror image of the claim to evidence check used for papers. What the mode deliberately does not do is encode the rules that actually get proposals rejected on format. The README says layer 6 distills only the stable structure of NSF and NIH proposals, and that for binding requirements such as page limits, formatting and deadlines you should consult the source: the NSF Proposal and Award Policies and Procedures Guide, the CAREER program page, and NIH's Write Your Application guide with its Specific Aims, Significance, Innovation and Approach sections. Hardcoding a page limit into a prompt would make it wrong within a funding cycle, so the prompt points outward instead.

The rules trim intensifiers and refuse to touch numbers

Four capabilities are claimed, and the boundary between them is precise. The first trims stock phrasing, naming openers about recent years, claims that something paves the way, references to extensive experiments, the hedge about being to the best of our knowledge, rule-of-three constructions, very long sentences and em-dashes. The second keeps claims tied to evidence by refusing any verb stronger than the data, so prove becomes show empirically, and by turning vague magnitudes into attributed ranges. The third leaves real scholarship alone, keeping evidence-tied hedging, passive voice where it fits, the word we, definitions, symbols and every citation, and it states that it does not change a number or a reference. The worked example in the README shows a stock opener followed by a contrast clause replaced by a sentence that names the actual limitation, then the three fronts of the work, then the two domains it is demonstrated on. Longer passes covering a general example, an NIH Specific Aims page and a funded NSF CAREER summary sit in examples/before-after.md.

The homepage is a vendor, and the license metadata is silent

Two housekeeping facts are worth knowing. The project's recorded homepage is https://morphmind.ai, which is one of the three agents the skill is said to run on rather than a page describing the skill itself. And the licensing is stated three ways that do not quite line up: the repository's license metadata records no standard identifier, the README states MIT, and a LICENSE file sits at the root. The reused catalog from the other project is separately credited as MIT. There are no GitHub releases here either, so there is no tag to pin, and the last commit on the default branch is dated 2026-07-03. The usage block assumes a LaTeX workflow, since the agent is pointed at a main.tex file and can optionally be told to match a voice from a prior paper PDF with a target venue named, ICLR in the example given.

Editorial conclusion

academic-humanizer is worth reading if you draft scholarly text with an AI assistant and dislike the generic register it produces, and it is cheap to try because the entire thing is one prompt file you can fork. Set your expectations from what it declines to do. It does not generate findings, change citations, or touch numbers, and its authors state plainly that it is not built to defeat AI-use detection, so it cannot help anyone avoid a disclosure obligation. Its rules encode one group's voice, which means the output quality depends on whether you calibrate it against your own accepted writing. For grant work, follow the funder's own guidance for page limits and formatting rather than trusting the skill, because it deliberately points at the source for those. And check the terms yourself: the repository's license metadata records nothing while the README states MIT.

Frequently asked questions

Official sources

1. AIScientists-Dev/academic-humanizer on GitHub
2. Issues
3. Project website
4. README