/humanize — AI-voice audit (detect-and-flag)
Read the target file (or all paper-like files), audit for the canonical AI-voice tells in academic prose, and write a structured report. The skill does not rewrite. The author edits.
Why this skill exists
Referees and editors increasingly recognise AI-generated prose. The tells are not stylistic preferences — they're statistically conspicuous patterns the LLM training distribution produces at higher rates than human academic writers. Five reasons to audit before submission:
- Reviewer suspicion is a tax. Even good substance pays a credibility tax if the prose reads as AI-drafted.
- Journal policy is tightening. A growing number of venues require disclosure or prohibit AI-drafted text.
- AI tells signal weak content. Boilerplate transitions ("Moreover", "It is important to note") almost always cover up logical gaps the author didn't think through.
- You are not the tells. Even authors who use AI tools heavily can preserve their own voice by stripping the model's lexical fingerprint.
- The fix is cheap once you can see it. The cost is detection, not rewriting — once the report flags the tells, removal is mechanical.
What this skill is NOT
- Not a rewriter. No
--rewrite mode. Auto-rewriting AI tells degrades prose quality (cross-vendor research finding); the author preserves voice by editing manually.
- Not a substance reviewer. Use
/review-paper for argument structure, identification, citations.
- Not a grammar checker. Use
/proofread for grammar, typos, overflow, citation format.
- Not a fact-checker. Use
/verify-claims for Chain-of-Verification fact-checking of citations and numeric claims.
/humanize is the voice lens. Run it alongside the others — none of them substitute.
When to use
- Before journal submission.
- Before posting a working paper / preprint / SSRN draft.
- After any AI-assisted prose generation (R&R response drafts, lit-review synthesis, abstract revisions).
- As a self-discipline pass after long writing sessions — your own writing drifts toward LLM patterns when you stare at LLM output all day.
When NOT to use
- On
.bib, .R, or other non-prose files — the detectors are tuned for academic prose.
- On code comments — the tells are different.
- On UI/UX copy — voice norms diverge.
Detection categories
The humanize-auditor agent checks these category groups:
1. BOILERPLATE TRANSITIONS
High-confidence AI tells when they appear sentence-initial or mid-paragraph as connective tissue:
Moreover, / Furthermore, / Additionally, / In addition,
It is important to note that / It is worth noting that / Notably,
In conclusion, / In summary, / To summarise,
On the other hand, (when not contrasting two named things)
Building on this, / Building upon this,
As we can see, / As is evident, / Indeed, (stacked)
Severity: HIGH if more than 1 per 1000 words. MED if 1 per 2000 words. LOW if rare but present.
2. AI-CLICHÉ LEXICON
Words and phrases statistically over-represented in LLM output relative to academic prose:
- "navigate the complexities", "navigate the landscape"
- "delve into", "delve deeper into"
- "tapestry of", "rich tapestry"
- "robust framework", "comprehensive framework", "holistic framework"
- "comprehensive approach" / "multifaceted approach" / "nuanced approach" (especially when stacked)
- "leverage" (as a verb in non-finance / non-engineering contexts)
- "in today's [X] landscape" / "in today's rapidly evolving"
- "play a crucial role" / "play a pivotal role" / "play a significant role"
- "shed light on"
- "underscore the importance" / "highlight the importance"
- "It is essential to" / "It is crucial to"
Severity: HIGH on a paper's first three pages (abstract, intro). MED elsewhere.
3. EM-DASH AND PUNCTUATION OVERUSE
- Em-dash overuse — more than 3 em-dashes per paragraph is a tell.
- Semicolon stacks — three or more semicolons in a single paragraph.
- Triple-Oxford-comma constructions — lists of three with deliberate parallelism repeated paragraph-to-paragraph.
Severity: MED. Em-dashes are a legitimate authorial choice; flag overuse, not all use.
4. SYMMETRIC PARAGRAPH SHAPES
Paragraphs with the same micro-architecture: topic sentence → three examples → summarising clause. Repeated across consecutive paragraphs is the AI tell — not the shape itself.
Detection: flag any three-paragraph window where each paragraph fits the topic→examples→summary cadence.
Severity: MED if 3-paragraph window; HIGH if 5+ paragraph stretch.
5. TRICOLON ABUSE
"X, Y, and Z" three-element lists are a legitimate rhetorical device. Tells are:
- More than 4 tricolons per page.
- Tricolons used for items that could naturally be 2 or 4.
- Adjective tricolons stacked ("clear, concise, and compelling"; "rigorous, robust, and reliable").
Severity: LOW if rare; MED if patterned.
6. HEDGING STACKING
Stacked epistemic hedges in single sentences:
- "might potentially be argued"
- "could possibly suggest"
- "may arguably"
- "perhaps potentially"
Severity: HIGH — these are almost never authorial choices; they're LLM uncertainty-management.
7. "NOT ONLY X, BUT ALSO Y" FRAMES
Used sparingly, this is a legitimate construction. AI tells:
- More than 2 per paper.
- Used when X and Y are not actually parallel.
- Used as paragraph openers.
Severity: MED.
8. FORMULAIC OPENERS
- Section openers of the form "This [paper / chapter / section / analysis] [does X]."
- Paragraph openers that re-state the section title.
- Abstract opening with "In this paper, we..." (legitimate in some sub-fields; flag for review where it's atypical, e.g., AER abstracts rarely use it).
Severity: LOW unless every section starts this way.
9. HYPHENATION EXCESS
Long chains of compound modifiers as a paragraph signature:
- "data-driven", "evidence-based", "well-suited", "well-established", "long-standing" — fine individually; flag if three or more appear in a single paragraph.
Severity: LOW.
10. SYCOPHANCY / SELF-IMPORTANT FRAMING
- "This important contribution"
- "This significant finding"
- "Our novel approach"
- Self-citation as "groundbreaking" / "pioneering"
Severity: HIGH — these read as AI-generated promotional copy; referees will react badly.
Steps
Identify files to audit:
- If
$ARGUMENTS starts with a filename: audit that file only.
- If
$ARGUMENTS is all: audit all .qmd, .tex, .md files in Slides/, Quarto/, root, and master_supporting_docs/.
- Skip
.bib, .R, .py, code files, and any file under scripts/.
Parse --severity flag (default: report all).
--severity low → report all findings.
--severity med → suppress LOW findings.
--severity high → report only HIGH findings.
For each file, launch the humanize-auditor agent with the 10 detection categories.
Receive structured report from the agent. Format per finding:
line N | category | severity | current text | suggested rewrite or "remove"
Write report to quality_reports/audits/humanize_<filename>_report.md. Include:
- Per-category counts (HIGH / MED / LOW)
- Per-finding table
- Summary recommendation (rough thresholds):
- > 8 HIGH findings per 1000 words: prose reads as AI-drafted. Author should rewrite the affected sections, not patch.
- 5–8 HIGH per 1000 words: substantial AI voice. Strip the tells before submission.
- < 5 HIGH per 1000 words: light cleanup; mostly cosmetic.
Present summary to user:
- Total findings per category
- Most concentrated paragraphs (top 3)
- Action recommendation (rewrite vs. strip vs. cosmetic)
Pairings
| Situation |
Do |
| When you've drafted prose with AI assistance |
Run /humanize before submission. Pair with /proofread (grammar) and /verify-claims (citations). |
| When you wrote in your own voice |
Run /humanize anyway — your own prose drifts toward LLM patterns after long sessions of AI-assisted work. |
| Submission-ready review |
/review-paper --peer [journal] --variance 3 for substance, /humanize for voice, /verify-claims for facts. |
Anti-pattern: no --rewrite mode
We deliberately do not ship /humanize --rewrite. Cross-vendor research (Cursor / Aider community findings; cited in the v1.9.0 plan) finds that auto-rewriting prose to strip AI tells degrades quality more often than it improves it — the rewriter introduces its own AI tells. The detect-and-flag pattern preserves authorial voice; the cost is your editing time, which is exactly the cost we want to pay.
If you find yourself reaching for an auto-rewriter, that's the signal to rewrite the paragraph from scratch — not to patch the tells one by one.
Output
- Report at
quality_reports/audits/humanize_<filename>_report.md (that subdirectory is gitignored).
- Summary to the conversation: counts per category, top concentrated paragraphs, action recommendation.
- No file edits. The user reads the report and applies changes manually.
Respect a documented voice profile
If voice-profile.md exists at the repo root, read it first. A habit the author has
declared deliberate — frequent em-dashes, first person, a particular connective — is not a
finding. Flagging a documented preference as an AI tell is a false positive, and false
positives erode the report's authority faster than misses do.
Build one with /voice-profile. This skill says what to remove;
that one says what to write toward.
What this skill cannot do (v2.5)
/humanize finds surface tells — boilerplate transitions, the AI-cliché lexicon, hedging
stacks, symmetric paragraph shapes. Fixing them improves readability, which is worth doing
whoever wrote the text.
It does not make prose stop reading as machine-generated to a detector. An article polished
through several rounds of surface de-AI-ing was submitted to Pangram, a neural AI-text
detector, and came back 100% AI-written. Those detectors classify on the token-level
statistics of LLM generation, which survive any transformation the model applies — because
every transformation is still LLM-generated text.
So: a clean report here means the prose reads well. It does not mean it reads human. If
provenance matters, the author writes the load-bearing sentences and measures with a real
detector. See writing-with-ai.md.
1---2name: humanize3description: Read-only audit of `.tex`, `.qmd`, or `.md` text for AI-voice tells — boilerplate transitions ("Moreover", "Furthermore", "It is important to note that"), AI-cliché lexicon ("delve", "navigate the complexities", "tapestry", "robust framework"), em-dash overuse, symmetric paragraph shapes, tricolon abuse, hedging stacking, "not only X but also Y" frames, and formulaic openers. Produces a report; does NOT rewrite. Use when user says "humanize", "does this sound like AI?", "check for AI tells", "de-AI this draft", "remove AI voice", "audit my prose for sycophancy", or before journal submission / posting a working paper.4argument-hint: "[filename or 'all'] [--severity low|med|high]"5disable-model-invocation: true6allowed-tools: ["Read", "Grep", "Glob", "Write", "Agent", "Task"]7disallowed-tools: ["Edit", "MultiEdit"]8metadata:9 source: https://skillmd.com/skills/pedrohcgs/humanize10 author: Claude Code Academic Workflow11 version: 1.0.012---1314# `/humanize` — AI-voice audit (detect-and-flag)1516Read the target file (or all paper-like files), audit for the canonical AI-voice tells in academic prose, and write a structured report. **The skill does not rewrite.** The author edits.1718## Why this skill exists1920Referees and editors increasingly recognise AI-generated prose. The tells are not stylistic preferences — they're statistically conspicuous patterns the LLM training distribution produces at higher rates than human academic writers. Five reasons to audit before submission:21221. **Reviewer suspicion is a tax.** Even good substance pays a credibility tax if the prose reads as AI-drafted.232. **Journal policy is tightening.** A growing number of venues require disclosure or prohibit AI-drafted text.243. **AI tells signal weak content.** Boilerplate transitions ("Moreover", "It is important to note") almost always cover up logical gaps the author didn't think through.254. **You are not the tells.** Even authors who use AI tools heavily can preserve their own voice by stripping the model's lexical fingerprint.265. **The fix is cheap once you can see it.** The cost is detection, not rewriting — once the report flags the tells, removal is mechanical.2728## What this skill is NOT2930- **Not a rewriter.** No `--rewrite` mode. Auto-rewriting AI tells degrades prose quality (cross-vendor research finding); the author preserves voice by editing manually.31- **Not a substance reviewer.** Use `/review-paper` for argument structure, identification, citations.32- **Not a grammar checker.** Use `/proofread` for grammar, typos, overflow, citation format.33- **Not a fact-checker.** Use `/verify-claims` for Chain-of-Verification fact-checking of citations and numeric claims.3435`/humanize` is the *voice* lens. Run it alongside the others — none of them substitute.3637## When to use3839- Before journal submission.40- Before posting a working paper / preprint / SSRN draft.41- After any AI-assisted prose generation (R&R response drafts, lit-review synthesis, abstract revisions).42- As a self-discipline pass after long writing sessions — your own writing drifts toward LLM patterns when you stare at LLM output all day.4344## When NOT to use4546- On `.bib`, `.R`, or other non-prose files — the detectors are tuned for academic prose.47- On code comments — the tells are different.48- On UI/UX copy — voice norms diverge.4950## Detection categories5152The humanize-auditor agent checks these category groups:5354### 1. BOILERPLATE TRANSITIONS5556High-confidence AI tells when they appear sentence-initial or mid-paragraph as connective tissue:5758- `Moreover,` / `Furthermore,` / `Additionally,` / `In addition,`59- `It is important to note that` / `It is worth noting that` / `Notably,`60- `In conclusion,` / `In summary,` / `To summarise,`61- `On the other hand,` (when not contrasting two named things)62- `Building on this,` / `Building upon this,`63- `As we can see,` / `As is evident,` / `Indeed,` (stacked)6465**Severity:** HIGH if more than 1 per 1000 words. MED if 1 per 2000 words. LOW if rare but present.6667### 2. AI-CLICHÉ LEXICON6869Words and phrases statistically over-represented in LLM output relative to academic prose:7071- "navigate the complexities", "navigate the landscape"72- "delve into", "delve deeper into"73- "tapestry of", "rich tapestry"74- "robust framework", "comprehensive framework", "holistic framework"75- "comprehensive approach" / "multifaceted approach" / "nuanced approach" (especially when stacked)76- "leverage" (as a verb in non-finance / non-engineering contexts)77- "in today's [X] landscape" / "in today's rapidly evolving"78- "play a crucial role" / "play a pivotal role" / "play a significant role"79- "shed light on"80- "underscore the importance" / "highlight the importance"81- "It is essential to" / "It is crucial to"8283**Severity:** HIGH on a paper's first three pages (abstract, intro). MED elsewhere.8485### 3. EM-DASH AND PUNCTUATION OVERUSE8687- Em-dash overuse — more than 3 em-dashes per paragraph is a tell.88- Semicolon stacks — three or more semicolons in a single paragraph.89- Triple-Oxford-comma constructions — lists of three with deliberate parallelism repeated paragraph-to-paragraph.9091**Severity:** MED. Em-dashes are a legitimate authorial choice; flag overuse, not all use.9293### 4. SYMMETRIC PARAGRAPH SHAPES9495Paragraphs with the same micro-architecture: topic sentence → three examples → summarising clause. Repeated across consecutive paragraphs is the AI tell — not the shape itself.9697**Detection:** flag any three-paragraph window where each paragraph fits the topic→examples→summary cadence.9899**Severity:** MED if 3-paragraph window; HIGH if 5+ paragraph stretch.100101### 5. TRICOLON ABUSE102103"X, Y, and Z" three-element lists are a legitimate rhetorical device. Tells are:104105- More than 4 tricolons per page.106- Tricolons used for items that could naturally be 2 or 4.107- Adjective tricolons stacked ("clear, concise, and compelling"; "rigorous, robust, and reliable").108109**Severity:** LOW if rare; MED if patterned.110111### 6. HEDGING STACKING112113Stacked epistemic hedges in single sentences:114115- "might potentially be argued"116- "could possibly suggest"117- "may arguably"118- "perhaps potentially"119120**Severity:** HIGH — these are almost never authorial choices; they're LLM uncertainty-management.121122### 7. "NOT ONLY X, BUT ALSO Y" FRAMES123124Used sparingly, this is a legitimate construction. AI tells:125126- More than 2 per paper.127- Used when X and Y are not actually parallel.128- Used as paragraph openers.129130**Severity:** MED.131132### 8. FORMULAIC OPENERS133134- Section openers of the form "This [paper / chapter / section / analysis] [does X]."135- Paragraph openers that re-state the section title.136- Abstract opening with "In this paper, we..." (legitimate in some sub-fields; flag for review where it's atypical, e.g., AER abstracts rarely use it).137138**Severity:** LOW unless every section starts this way.139140### 9. HYPHENATION EXCESS141142Long chains of compound modifiers as a paragraph signature:143144- "data-driven", "evidence-based", "well-suited", "well-established", "long-standing" — fine individually; flag if three or more appear in a single paragraph.145146**Severity:** LOW.147148### 10. SYCOPHANCY / SELF-IMPORTANT FRAMING149150- "This important contribution"151- "This significant finding"152- "Our novel approach"153- Self-citation as "groundbreaking" / "pioneering"154155**Severity:** HIGH — these read as AI-generated promotional copy; referees will react badly.156157## Steps1581591. **Identify files to audit:**160 - If `$ARGUMENTS` starts with a filename: audit that file only.161 - If `$ARGUMENTS` is `all`: audit all `.qmd`, `.tex`, `.md` files in `Slides/`, `Quarto/`, root, and `master_supporting_docs/`.162 - Skip `.bib`, `.R`, `.py`, code files, and any file under `scripts/`.1631642. **Parse `--severity` flag** (default: report all).165 - `--severity low` → report all findings.166 - `--severity med` → suppress LOW findings.167 - `--severity high` → report only HIGH findings.1681693. **For each file, launch the `humanize-auditor` agent** with the 10 detection categories.1701714. **Receive structured report** from the agent. Format per finding:172173 ```174 line N | category | severity | current text | suggested rewrite or "remove"175 ```1761775. **Write report** to `quality_reports/audits/humanize_<filename>_report.md`. Include:178 - Per-category counts (HIGH / MED / LOW)179 - Per-finding table180 - Summary recommendation (rough thresholds):181 - **> 8 HIGH findings per 1000 words**: prose reads as AI-drafted. Author should rewrite the affected sections, not patch.182 - **5–8 HIGH per 1000 words**: substantial AI voice. Strip the tells before submission.183 - **< 5 HIGH per 1000 words**: light cleanup; mostly cosmetic.1841856. **Present summary** to user:186 - Total findings per category187 - Most concentrated paragraphs (top 3)188 - Action recommendation (rewrite vs. strip vs. cosmetic)189190## Pairings191192| Situation | Do |193|---|---|194| When you've drafted prose with AI assistance | Run `/humanize` before submission. Pair with `/proofread` (grammar) and `/verify-claims` (citations). |195| When you wrote in your own voice | Run `/humanize` anyway — your own prose drifts toward LLM patterns after long sessions of AI-assisted work. |196| Submission-ready review | `/review-paper --peer [journal] --variance 3` for substance, `/humanize` for voice, `/verify-claims` for facts. |197198## Anti-pattern: no `--rewrite` mode199200We deliberately do not ship `/humanize --rewrite`. Cross-vendor research (Cursor / Aider community findings; cited in the v1.9.0 plan) finds that auto-rewriting prose to strip AI tells degrades quality more often than it improves it — the rewriter introduces its *own* AI tells. The detect-and-flag pattern preserves authorial voice; the cost is your editing time, which is exactly the cost we want to pay.201202If you find yourself reaching for an auto-rewriter, that's the signal to rewrite the paragraph from scratch — not to patch the tells one by one.203204## Output205206- Report at `quality_reports/audits/humanize_<filename>_report.md` (that subdirectory is gitignored).207- Summary to the conversation: counts per category, top concentrated paragraphs, action recommendation.208- **No file edits.** The user reads the report and applies changes manually.209210## Respect a documented voice profile211212If `voice-profile.md` exists at the repo root, **read it first**. A habit the author has213declared deliberate — frequent em-dashes, first person, a particular connective — is **not a214finding**. Flagging a documented preference as an AI tell is a false positive, and false215positives erode the report's authority faster than misses do.216217Build one with [`/voice-profile`](../voice-profile/SKILL.md). This skill says what to remove;218that one says what to write toward.219220## What this skill cannot do (v2.5)221222`/humanize` finds **surface tells** — boilerplate transitions, the AI-cliché lexicon, hedging223stacks, symmetric paragraph shapes. Fixing them improves readability, which is worth doing224whoever wrote the text.225226**It does not make prose stop reading as machine-generated to a detector.** An article polished227through several rounds of surface de-AI-ing was submitted to Pangram, a neural AI-text228detector, and came back **100% AI-written**. Those detectors classify on the token-level229statistics of LLM generation, which survive any transformation the model applies — because230every transformation is still LLM-generated text.231232So: **a clean report here means the prose reads well. It does not mean it reads human.** If233provenance matters, the author writes the load-bearing sentences and measures with a real234detector. See [`writing-with-ai.md`](../../rules/writing-with-ai.md).