Analyzes a draft for the patterns that make writing read as AI-generated — across structure, vocabulary, transitions, rhythm, punctuation, and closings — and returns a scored authenticity audit with a specific rewrite directive for each defect, rewriting toward the author's real voice rather than mechanically deleting patterns. Use whenever someone wants to humanize text, make writing sound less robotic or less like ChatGPT, remove AI tells, check whether a draft "sounds AI," or de-slop an AI-assisted newsletter, social, or long-form draft before publishing. Reads a VOICE-PRINT.md when present so rewrites have a target. Run this for how a draft reads; run content-review for whether it should ship. Not a generic grammar/copyedit pass and not a find-and-replace tool.
Analyzes a draft for the patterns that make writing read as AI-generated — across structure, vocabulary, transitions, rhythm, punctuation, and closings — and returns a scored authenticity audit with a specific rewrite directive for each defect, rewriting toward the author's real voice rather than mechanically deleting patterns. Use whenever someone wants to humanize text, make writing sound less robotic or less like ChatGPT, remove AI tells, check whether a draft "sounds AI," or de-slop an AI-assisted newsletter, social, or long-form draft before publishing. Reads a VOICE-PRINT.md when present so rewrites have a target. Run this for how a draft reads; run content-review for whether it should ship. Not a generic grammar/copyedit pass and not a find-and-replace tool.
triggers
["/ai-slop-audit","/humanize","user wants to humanize text or make writing sound less robotic","user asks \"does this sound like AI\" / \"remove the AI tells\"","de-slop a draft before publishing","reviewing AI-assisted content for authenticity"]
This skill does two jobs the way a good editor does: it names what makes a draft read as machine-written, and it rewrites each defect toward a real human voice — not by deleting patterns mechanically, but by replacing them with the specific, evidence-bearing, rhythmically-varied writing that AI text lacks.
It is the analysis half of "analyze the most common defects of AI-generated content, and provide the key content-creation reference." The taxonomy is the analysis; the per-defect rewrite directives are the reference.
The governing distinction: an AI tell is a pattern and a frequency, almost never a single token. Em-dashes are good writing; the tell is using them as the only rhythmic device. The rule of three is good rhetoric; the tell is using it in every paragraph. This skill therefore scores density and co-occurrence, and it never recommends a blanket ban on a punctuation mark or a word. Mechanical de-slopping (delete all em-dashes, swap every "leverage" for "use") produces stiff, lobotomized text — a different kind of bad writing. The fix is always: replace the pattern with something specific to this author and this point.
Routing — this skill vs. content-review
These two overlap on voice, so route by what the request is about:
Run this, not content-review, when the request is about how the writing reads — "sound less like AI," "humanize this," "does this sound like ChatGPT," "remove the tells," "make it sound like a person."
Run content-review instead when the request is about whether it should ship — ICP fit, funnel stage, CTA, differentiation, platform format, launch-readiness scoring.
They compose: run ai-slop-audit first to fix authenticity, then content-review to score readiness.
The two share one banned-hype canon: content-review owns that table as the single source of truth; this skill references it (see taxonomy family 2.1) rather than maintaining a second copy. This skill's authenticity verdict feeds content-review's brand-voice dimension (Dimension 6) — a draft that fails here cannot pass that check.
Inputs
Input
Required?
Description
The draft
Required
The text to audit. Any length; note if it's an excerpt.
VOICE-PRINT.md
Strongly recommended
Output of voice-print. Gives the rewrite a target instead of a generic one. Without it, rewrites aim at a generic "specific, honest practitioner" voice and the skill says so.
Channel
Optional
Newsletter / LinkedIn / X / blog — calibrates which structural tells matter most (e.g. listicle skeleton is worse in an essay than in a carousel).
Author intent
Optional
What the piece is trying to do. Prevents flagging a deliberate choice as a defect.
BLOCK rule
IF no draft text is provided:
BLOCK. Return: "Paste the draft you want audited for AI tells. If you have a
VOICE-PRINT.md, include it — the rewrites will aim at that voice instead of a generic one."
Note on origin: if the user says the text is not AI-written (it's their own), still run the audit — humans absorb these patterns too, especially the hype/transition vocabulary. Frame findings as "reads as AI" not "is AI," and weight the voice print heavily.
Decision Logic
Step 1 — Scan the six defect families
The full taxonomy — every pattern, why it reads as AI, and the rewrite directive — is in references/ai-tell-taxonomy.md. Read it before auditing. The six families:
#
Family
The tell in one line
1
Structural
Revelation hook ("Here's what no one tells you"), "it's not X, it's Y" reduction, symmetrical listicle skeleton, manufactured stakes
2
Lexical
Hype words (leverage, unlock, seamless, robust), filler ("it's worth noting"), AI-signature words (delve, tapestry, realm, navigate, ever-evolving)
3
Connective
Transition-word scaffolding (moreover, furthermore, additionally, consequently) doing the work real logic should do
4
Rhythm
Uniform sentence length, equal paragraph blocks, relentless parallelism, no fragments, no variance
5
Punctuation/format
The em-dash-as-only-tool tic, bold-everything, emoji bullets, title-case headers, list-for-everything
6
Closing
"In conclusion / Ultimately / At the end of the day," recap-summary, tacked-on uplift or CTA
Step 2 — Score by density, not presence
For each family, assess severity on how pervasive it is, not whether it appears once:
none — not present, or a single deliberate instance that works
light — present but occasional; doesn't dominate the read
moderate — recurring; a reader would start to notice the pattern
heavy — the pattern is load-bearing; remove it and the structure collapses
Then read co-occurrence: the unmistakable AI signature is several families heavy at once — revelation hook + transition scaffolding + uniform rhythm + em-dash tic + "in conclusion." One family heavy = a tic to fix. Four families heavy = the text has no voice; rewrite, don't patch.
Step 3 — Distinguish tell from legitimate use (the false-positive gate)
Before flagging, apply the gate for that pattern (detailed per-pattern in the taxonomy). The general test:
Frequency: is it the only tool, or one of several? (Em-dashes once a paragraph = fine.)
Function: does it carry real meaning, or substitute for it? ("It's not X, it's Y" stating a genuine reframe = fine; doing it three times for rhythm = tell.)
Voice fit: does the VOICE-PRINT.md record this as a genuine signature move? If the author really does open on a bold claim, that's their voice, not slop.
Flag only what fails the gate. Over-flagging is its own failure — it trains the user to strip out real writing.
Step 4 — Rewrite toward the voice, per defect
For each flagged item, produce a rewrite directive, not just a deletion:
Quote the offending line.
Name the family + why it reads as AI.
Give a specific rewrite that does what the line was trying to do — using the author's voice (from the voice print) and real specificity (a number, a scene, a named thing). If no voice print, aim at the practitioner standard in pmm/DOMAIN.md and say the rewrite is generic.
The reframing principle (cross-ref content-craft): replace the function the slop was faking. A revelation hook fakes "this is worth your attention" — replace it with an actual specific stake. Transition words fake logical connection — replace with the real causal link or cut.
Step 5 — Verdict
The deciding variable between DE-SLOP and REWRITE is does a genuine voice anchor survive? — is there at least one contiguous passage that is really the author, not the machine. Family count alone doesn't decide it.
CLEAN — no family above light; reads as a real person. Optionally note 1 polish.
DE-SLOP — tells are present (1–3 families moderate/heavy) AND a voice anchor
survives. Provide rewrite directives. When slop dominates but one real
passage holds, the directive is "build outward from the anchor, cut the
rest" — not patch-in-place.
REWRITE — heavy tells across many families (typically 4+) AND no passage survives
the gate; there's no voice anchor anywhere. Recommend: extract the one
real idea, run voice-print, then redraft. Patching won't save it.
Outputs
## AI-Slop Audit
**Draft:** [title/desc] **Channel:** [if given] **Voice print:** [used / none — rewrites generic]
### Signature read
[1–2 sentences: does this read as AI, and why — name the co-occurring families.]
### Defect scan
| Family | Severity | What's driving it (quoted) |
| Structural | [none/light/moderate/heavy] | "..." |
| ... one row per family |
### Rewrite directives
[Per flagged item:]
**[Family] — "[quoted line]"**
- Why it reads as AI: [one line]
- Rewrite: [specific replacement in the author's voice, with real specificity]
### Verdict: [CLEAN / DE-SLOP / REWRITE]
[Rationale. If REWRITE: point to voice-print + redraft.]
### What's already human
[Lines that work — preserve these. Never omit; de-slopping must protect real voice.]
Feeds: the verdict and brand-voice findings flow into content-review Dimension 6 (Brand voice). A draft that fails ai-slop-audit cannot pass content-review's brand-voice check.
Cross-Review Triggers
This is itself a review skill; it does not wait on a review of its own. It readsvoice-print output and feedscontent-review (its authenticity verdict is consumed by content-review's brand-voice dimension — that's the blocking_reviews direction in skill.meta.json: content-review's brand-voice check depends on this, not the reverse). When customer-facing copy is in scope, run ai-slop-audit before content-review so authenticity is fixed before the full GTM scoring.
Example Usage
Example 1 — Classic AI draft, no voice print
Input: A LinkedIn post opening "Here's the truth no one talks about: AI isn't replacing developers, it's amplifying them. Moreover, the paradigm shift we're seeing is seamless."
Output: Signature read = "reads as AI: revelation hook + 'it's not X it's Y' + transition scaffolding + hype all co-occurring." Verdict DE-SLOP with rewrites; flags it has no voice print and recommends one for a sharper rewrite.
Example 2 — Human draft with absorbed tics
Input: A founder's real essay that nonetheless leans on "it's worth noting," "furthermore," and ends "Ultimately, the future is bright." + a VOICE-PRINT.md.
Output: CLEAN-leaning-DE-SLOP: structural/rhythm = light (it's genuinely theirs), but connective + closing = moderate. Rewrites the transitions and the ending toward their documented closing instinct (worldview, not uplift). Preserves the rest.
Example 3 — Over-correction guard
Input: A draft with three well-used em-dashes and one rule-of-three that lands.
Output: CLEAN. Explicitly does not flag the em-dashes or the triple — they pass the frequency/function gate. Notes that removing them would harm the writing.
Validation Criteria
Output passes if:
Severity is scored by density/co-occurrence, not mere presence
The false-positive gate is applied — deliberate, well-functioning uses are NOT flagged
Every flagged item has a specific rewrite directive, not just "remove this"
Rewrites aim at the voice print when present; when absent, the output says rewrites are generic
A "what's already human" section preserves real voice
Verdict follows the CLEAN/DE-SLOP/REWRITE logic; REWRITE points back to voice-print
No recommendation to blanket-ban a punctuation mark or single word
Output fails if:
It flags every em-dash / every "leverage" mechanically (find-replace behavior)
Rewrites are generic deletions that leave stiff text
It treats one instance of a pattern as proof of AI authorship
It strips a documented signature move because it superficially resembles a tell
Benchmarks
Authenticity is partly subjective, so this skill is validated primarily by golden example. Two checkable signals do anchor it:
