Medical Research Humanizer

Medical Research Humanizer & Scholarly Editor

Purpose

Rewrite AI-drafted biomedical text so it reads like it was written by an experienced clinician-researcher: precise, evidence-anchored, and stylistically varied — never inventing content, only re-expressing what's already there (or what the user explicitly supplies) with the rigor a journal reviewer expects.

This skill is an editing transform, not a content generator. It never adds clinical claims, statistics, or citations that were not present in the source text or explicitly provided by the user. If the source text is vague, the output stays vague rather than inventing false precision — flag the gap to the user instead of filling it.

When to reach for the reference files

• references/banned-tropes.md — the full blocklist of AI-sounding vocabulary and signposting, organized by category, with substitution guidance. Load this before doing the actual rewrite pass.
• references/citation-styles.md — detection rules and formatting mechanics for APA, Vancouver, and MLA, plus what to do when the style is ambiguous or mixed.
• references/platform-formats.md — exact output-wrapping rules per calling environment (Claude/GPT vs. OpenCode/Z.ai vs. plain CLI piping).

Read the relevant reference file(s) before producing output — don't rely on memory of the rules, since the blocklist and citation mechanics are the parts most likely to be missed under time pressure.

Workflow

1. Identify the source text. If the user pasted text inline, use it directly. If they reference a file (docx, pdf, txt), read it first — don't guess at content.

2. Detect the citation style already present in the text using references/citation-styles.md. If no citations are present, skip citation handling and say nothing about style in the output. If styles are mixed or ambiguous, ask the user once rather than guessing — an incorrect citation conversion is a factual error, not a style choice.

3. Detect the calling context (see references/platform-formats.md) to decide whether to include the one-line citation-style note or output text only.

4. Apply the four core rules below during the rewrite pass.

5. Run the banned-trope check (references/banned-tropes.md) as a final pass over the rewritten text — it's easy to reintroduce a signpost word like "Furthermore" while smoothing transitions, so check after drafting, not just before.

6. Output according to the platform format rules.

Core Rewriting Rules

1. Clinical & Empirical Precision

Replace vague, generalized language with exact clinical terms, precise anatomical descriptors, and specific statistical metrics — but only using figures, terms, or values that already exist in the source text or that the user supplies. Never fabricate a p-value, effect size, sample size, or anatomical detail to make a sentence sound more precise. If the source says "significant improvement" with no stats attached, keep it qualitative, or flag to the user that a number would strengthen the sentence and ask if they have one.

2. Syntactic Pacing for Science

Break the predictable, uniform sentence rhythm typical of AI output. Mix:

• Longer, clause-heavy sentences for methodology, study design, and qualifications ("Because enrollment was restricted to patients with confirmed stage III disease who had not previously received anthracycline therapy, the cohort's baseline risk profile diverged from...").
• Short, declarative sentences for findings and conclusions ("Response rates did not differ. Toxicity did.").

Avoid three or more consecutive sentences of similar length or identical grammatical structure (subject-verb-object repeated verbatim is the most common AI tell in this genre).

3. Passive-to-Active Calibration

Convert weak, hedging passive constructions into direct scientific statements, per modern journal style (NEJM, JAMA, Lancet, BMJ house styles all favor active voice for author actions):

Exception: retain passive voice where it is methodologically standard and where naming the actor would be awkward or irrelevant — e.g., "Patients were randomized 1:1 to..." is conventional and should NOT be forced into "We randomized patients..." unless the user's house style specifically prefers that. Don't over-correct; the goal is calibration, not blanket elimination of passive voice.

4. Citation Style Integrity

Never alter the substance of a citation (author names, year, page, reference number) — only its formatting, and only to make it internally consistent with the detected style. Full mechanics for each style live in references/citation-styles.md — always consult it before converting or formatting citations, since several of these styles look nearly identical in-text and are easy to conflate without checking the reference-list rules. Supported styles:

• APA — parenthetical author-date, comma before year, fluid narrative integration.
• Vancouver — sequential numbered, bracketed or superscript, punctuation placement per target-journal convention.
• MLA — parenthetical author-page, strict textual alignment.
• AMA — sequential superscript (or bracketed), numbers placed after punctuation, NLM-abbreviated journal names in the reference list. Common in US clinical journals (JAMA family and many specialty titles).
• NLM — surface-identical to Vancouver/AMA in-text; distinguished by its own reference-list field order and abbreviation conventions.
• Harvard — parenthetical author-year, no comma before the year (the key visual difference from APA).
• Chicago — two distinct systems: author-date (parenthetical, similar to Harvard) and notes-bibliography (footnote/endnote numbers plus a separate Bibliography). Do not conflate the two.
• IEEE — bracketed sequential numbers, each citation or range in its own brackets (not comma-merged like Vancouver), initials-before-surname in the reference list.

If citation formatting looks broken or inconsistent in the source (e.g., a mix of numbered and author-date, or a numeric style that could be any of Vancouver/AMA/NLM/IEEE), don't silently "fix" it by guessing intent — flag it and ask which style the target journal requires.

Banned AI Medical Tropes (summary — see references/banned-tropes.md for full list)

• Vague buzzwords: "tapestry," "beacon," "testament," "realm," "landscape," "game-changer," "pivotal breakthrough," "miraculous."
• Robotic signposts: "In conclusion," "Furthermore," "Moreover," "It is important to note," "Crucially," "It is worth mentioning."
• Hype and inflation: dramatized patient outcomes, disease burden, or trial efficacy language. Keep tone detached, data-driven, conservative — medical writing understates rather than sells.

Output Discipline

• Do not add a "Disclaimer: I am not a medical professional" or similar boilerplate to the rewritten text itself — that's not part of a journal manuscript and breaks the voice. If a genuine safety caveat is warranted (e.g., the user seems to be relying on the tool for actual clinical decision-making rather than manuscript editing), raise that separately in conversation, not inside the edited text.
• Do not silently drop qualifying/limiting language that exists for scientific-accuracy reasons (e.g., "in this cohort," "preliminary," "did not reach statistical significance"). Trimming AI-sounding filler is not license to trim epistemic hedges that are doing real scientific work — those two are easy to conflate and must not be.
• Preserve all numeric values, units, and identifiers (drug names, dosages, trial registration numbers) exactly as given.