---
name: scholar
version: 2.0.0
description: |
  Remove signs of AI-generated writing from academic prose: journal articles,
  theses, dissertations, grant proposals, conference papers. Fixes the syntax
  tells (fused claims, semicolon splices, relative-clause chains, question-word
  subjects, bare demonstrative openers, vague summary nouns), the provenance
  tells (invented citations and invented technical specificity), and the claim
  tells (significance inflation, vague attribution, formulaic limitations).
  Matches the author's own published prose when a corpus is supplied. Preserves
  passive voice in methods, calibrated hedging, and sentence complexity. Based on Humanizer skill.
argument-hint: "[draft path] [corpus dir of the author's prior papers] [exemplar doc]"
license: MIT
allowed-tools:
  - Read
  - Write
  - Edit
  - Grep
  - Glob
  - Bash
---

# Academic Humanizer

Edit academic prose to remove AI tells without damaging what scholarly writing
depends on. 

---

## A. Provenance

**Describe standard methods in the field's standard words.** Fresh imagery
for a routine tool is a tell, because every specialist describes it the same
way and the author's own papers already contain that description. "The model is
accurate but unreadable, since it spreads the response across hundreds of
trees" → "Such models are often considered black boxes, because their
decision-making processes are not easily understood."

---

## B. Syntax

These four are what makes a technically correct paragraph still read as
machine-written. Authors report them as "weird logic" or "jumping around."

**One main claim per sentence.** The defect is not length. A 40-word
sentence with one claim and its conditions attached is normal scientific
prose. A 30-word sentence carrying three claims is not.

- **Front-load the condition.** "During X, Y occurred." Not "Y occurred, during
  periods when X, which is inconsistent with…"
- **No backtracking.** A clause that sends the reader to something already
  passed belongs earlier or in its own sentence.
- **Semicolons for listing only**, and only when the items contain commas. A
  semicolon joining two independent clauses leaves their relation unstated: the
  reader cannot tell whether the second elaborates the first, follows from it,
  or merely comes next. Name the relation with a conjunction, or use a period.
- **Cap coordination.** Three or more `and`-linked claims in one sentence is the
  real tell.

> **Before:** Nitrate was enhanced during late-morning periods of declining
> relative humidity, which is inconsistent with a response to prevailing
> humidity alone; relative humidity tendency and aerosol liquid water
> diagnostics indicate retention of particle water inherited from the preceding
> high-humidity period, allowing continued partitioning of HNO₃.
>
> **After:** During the late-morning RH decline, nitrate accumulation
> continued, which is inconsistent with a response to instantaneous RH
> magnitude alone. The RH tendency and ALWC diagnostics indicate that
> nitrate-rich mixed particles retain water from the preceding high-humidity
> period, providing a reservoir for continued HNO₃ partitioning.

Two sentences of ~26 words, both still complex, both still hedged. Split at the
claim boundary, then move the condition to the front. **Never fix this by
chopping into short declaratives** ("RH declines through the late morning.
Nitrate accumulates during that decline.") That reads as a lab report, and it
downgrades the epistemics: "may provide a reservoir" becomes "provides."

**One relative clause per sentence, none nested.** Watch `of which`,
`to which`, `in which`, `for which`, `against which`, `whose`, `each of which`.
Promote the inner clause to its own sentence or convert it to a prepositional
phrase.

> **Before:** Four responses are examined, each one for which nonlinearity is
> expected, and each of which is paired with independent evidence against which
> the relationship is tested.
>
> **After:** Four responses are examined. Each is one for which nonlinearity is
> expected from established chemistry, and each is paired with independent
> evidence.

**No question-word subjects.** Sentence-initial `What … is`, `How much …
depends`, `Which …`, `Whether … determines`. The sentence withholds its subject
and makes the reader carry an open question until the verb arrives. Name the
thing the question is about and put that noun in subject position.

| Question form | Noun form |
|---|---|
| How much forms depends on… | The amount formed depends on… |
| What is needed instead is X | X is required instead |
| Whether X responds determines Y | The existence of such a limit determines Y |
| What is measured at the receptor | The concentration measured at the receptor |

*Carve-out:* `whether` and `how` as verb complements are standard and stay
("examines whether nitrate follows RH history," "requires knowing how formation
responds"). The rule targets subject position only.

**The subject must name something.** Three failures:

- **Abstract process nouns.** "The attribution examined the response" has no
  referent in the world. Put the studied object in subject position, not the act
  of studying it. Watch: the attribution, the interpretation, the analysis, the
  examination of, the consideration of.
- **Bare demonstrative openers.** No paragraph or section may open with `This`,
  `That`, `These`, `Those`, or `It` as a bare subject: the referent sits in a
  paragraph the reader has left. `That response is nonlinear` → `Secondary
  pollutant formation responds nonlinearly to precursor and meteorological
  conditions`. A demonstrative *with a head noun* ("These policies," "This weak
  response") is fine mid-section. Deleting the opener and starting from the
  concrete claim is usually the best fix.
- **Vague summary nouns.** "the range," "the picture," "this dynamic," "the
  landscape," "the complexity." `Three cases illustrate the range` → the range
  of what? → `Particulate nitrate, ozone, and secondary organic aerosol
  illustrate three distinct forms`.

---

## C. Claims

**No significance inflation.** Watch: represents a significant advance,
paradigm shift, profound implications, opens new avenues, sheds light on,
fills a crucial gap, transformative, groundbreaking. State what you found and
what it does not establish. Let the reader assess significance.

**Name agents.** "Studies have shown" and "the literature suggests" cannot
be checked, and they flatten live disputes into fake consensus. Cite who found
what, and where the disagreement is.

**Match the verb to the evidence.** Cut dramatically, substantially,
markedly, clearly demonstrates, conclusively. A correlational design keeps
correlational verbs: predicts, is associated with. Never causes, drives, or
leads to unless the design licenses it.

**Specify methods, do not praise them.** "Robust," "comprehensive," and
"state-of-the-art" are not reproducible. Give the model, the software version,
the exclusions.

**Cut throat-clearing and manufactured gaps.** "In recent years there has
been growing interest" says nothing. "Little is known about X" is usually false
or trivially true. Open with the actual state of the question, and warrant your
study by what a specific prior design could not identify.

**Limitations must be specific.** Name what could be wrong and how it would
be detected, not "the sample was small, future research should use longitudinal
designs."

---

## D. Mechanics

**Zero em dashes in running prose.** Hard constraint. Restructure, never
substitute another dash. Paired aside → parentheses. Restatement → colon.
Contrast → new sentence. Caveat → "though," "because," "since." Trailing
afterthought → its own sentence, or cut it (never a semicolon; see B).
Unspaced en dashes in ranges (1994–2003) are correct and stay.

**Match the author's orthography.** Spelling locale from their corpus, not
your default: mixing `-ise`/`-ize` or `analyse`/`analyze` in one manuscript is a
visible copy-editing failure. Same for hyphenation of recurring compounds and
where abbreviations get expanded.

**Vocabulary watchlist.** delve into, underscore, highlight (verb), crucial,
pivotal, key (adj), landscape, interplay, intricate, nuanced, robust
(non-statistical), leverage, align with, foster, myriad, multifaceted, holistic,
navigate (figurative), realm, testament. Also trailing participles that assert
an inference ("…, highlighting the role of X"), forced triads, and "not merely
X but Y."

**Cut filler; keep one load-bearing hedge.** "In order to" → "To." "Due to
the fact that" → "Because." "It is important to note that" → delete. But "It
could potentially be argued that X might possibly suggest" → "These data are
consistent with X," which is *more* cautious, not less.

---

## What Not To Fix

- **Passive voice in methods and results.** The agent is irrelevant there.
- **Calibrated hedging.** "Suggests," "consistent with," "may indicate" carry
  epistemic content. Cut stacking, not hedging.
- **Sentence complexity.** 25 to 40 words with one claim and its conditions is
  the normal unit. Split at the claim boundary, never into short declaratives.
- **Hyphenated compound modifiers.** "Fixed-effects model" keeps its hyphen.
- **Repeated key terms.** Consistent terminology beats elegant variation; a new
  word implies a new construct.
- **Nominalizations that name constructs.** "Measurement invariance" is the name
  of the thing.
- **Structural signposting.** "Section 3 develops the model" is a reader aid.
- **Bold run-in headers when the template uses them.** Proposal forms and thesis
  plans often require them. The document's own format wins.

---

## Process

1. Identify venue and discipline. Ask if unclear.
2. Load the author's corpus first, before editing anything. Prior papers,
   drafts, lab publications. Extract their phrasing for recurring claims, their
   terminology, their spelling locale, their hedging register, their typical
   sentence shape. Where the draft restates something the author has already
   written, reuse their sentence. Generation is the fallback, not the default.
3. If an exemplar in the same genre is supplied (an approved proposal, an
   accepted thesis from the same department), count words section by section and
   report the comparison as a table. Do not cut on your own judgment.
4. Work through A and C, then D, then B. Syntax comes last, because it rewrites
   whole sentences and the content should be settled before that.
5. Check nothing in *What Not To Fix* was damaged, and that no claim got
   stronger.
6. Check structural consistency after restructuring: enumerations must match
   what follows ("three cases: X, Y, Z" above two paragraphs is the classic
   residue), and connectives must match actual position (a paragraph moved to
   the front cannot open with "also" or "third").
7. Grep and clear:
   - `;` outside a list or citation → *one claim per sentence*
   - paragraph-initial `This |That |These |Those |It ` → *subject must name
     something*
   - sentence-initial `What |How |Which |Whether ` → *no question-word subjects*
   - `of which|to which|in which|whose` → *one relative clause*
   - `—` → *zero em dashes*, count must be zero
   - `isation|ised|ising|analyse|behaviour|favour|modelling|levelled|centre|
     utilis` → *match the author's orthography*
8. Ask "what still reads as machine-generated here?", then revise once more.

**Output.** Editing a file: apply the edits, then report only the changes and
the flags. Returning prose in chat: final version, then the flags. Flags are
unverifiable citations, technical claims more specific than any traceable
source, and places where the original meaning was ambiguous and you chose.