De-AI your content
Applies to every AI-assisted draft before publishing or sending. Produces the same copy with the machine tells removed, the vague claims made concrete, and a full detection battery passed before it ships.
The economics: readers pattern-match AI tells in about a second and discount everything after, feeds suppress the formulaic shape, and answer engines never cite the median take. An unedited AI draft is the one piece of content guaranteed to look like everyone else's. This pass injects the only thing that differentiates, which is structure a machine would not default to and specifics a machine could not know.
The pass, in order
The order is the method. Reversing it produces a draft that has been word-swapped and still reads as generated.
X-ray the skeleton. Structural tells expose a draft even after every suspicious word is swapped, so structure gets fixed first. Scan against category 1 of the inventory and rewrite the sentence, never just the word.
Strip the surface. Run categories 2 through 5: punctuation, lexicon, content tells, voice tells.
Run the specificity pass. Every generic claim becomes a real number, name, or date pulled from source material, or it gets cut. If the source lacks the specific, flag the gap. Fabricated specificity is worse slop than vagueness, and in regulated industries it is a liability.
Run the rhythm pass. Read the piece aloud. Vary sentence length on purpose. Leave one fragment. Break at least one parallel structure per section. The target is copy that reads like one person typed it in one sitting.
Reconcile with the author. If a voice fingerprint document exists, load it and let it win every conflict with the tell list. Some humans genuinely open with rhetorical questions. Sanding off real habits produces a second kind of fake their audience catches just as fast.
Calibrate for the channel. Feed posts, cold email, long-form, and proof content each punish different tells.
Prove it clean. Run the full detection battery. A draft ships when it passes every applicable test, and never before — the correlation check is the one exception, since it only applies to teams running several accounts.
The reference pages
references/tell-inventory.md carries the full catalog in five categories, ordered by damage, with the fix for every entry. Structural findings mean rewriting sentences and merging paragraphs, and that happens before any surface fix, or the surface work gets overwritten.references/channel-calibration.md covers feed posts, cold email, newsletters and long-form, and case studies and proof content.references/detection-tests.md carries the seven tests, what each one catches, and what a repeated failure means.references/voice-fingerprint.md covers building the author document that overrides the tell list.references/worked-examples.md walks three passes end to end, plus one case where a flagged construction was correct and stayed.
The rule underneath the lexicon
Any phrase that appears in everyone's AI drafts belongs to no one's voice. That is the test to apply when a word is not on the list yet.
The same logic governs the specificity pass. A claim only a specific operator could make is the whole asset. A claim any competitor could publish unchanged argues for the category and says nothing about the company.
What good looks like
The best operator reads for skeleton first: flips, triads, uniform rhythm. The mediocre version swaps "delve," deletes the em dashes, and ships, leaving the skeleton intact. The skeleton is the tell.
The output is good when a reader who knows the author picks the draft out of a lineup at no better than chance, every specific traces to source material, at least one line costs the author something, and the author's actual quirks survived the pass.
When a draft keeps failing the lineup or the x-ray after repeated passes, the problem is upstream. The source material is too thin or the voice fingerprint is missing, and no amount of polish substitutes for either.
Rules
- MUST rewrite structure before touching word choice.
- MUST trace every number and name to source material, or cut it.
- MUST run every applicable detection test before anything ships (the correlation check applies only to multi-account teams).
- NEVER compensate with slang, forced edge, or manufactured casualness.
- NEVER treat paraphrase alone as humanizing. It preserves the structure that gives the draft away.
- NEVER override an author's real voice with the tell list.
- NEVER invent a specific to fill a gap. Flag the gap.
1---2name: content-de-slop-ai3description: Use this skill when an AI-assisted draft is about to publish or send: a blog post, a feed post, a newsletter, a landing page, a cold email, a case study, or a sequence step. Produces the same copy with the machine tells removed, the vague claims made concrete, and a detection battery passed before it ships. Covers de-slopping AI content, humanizing AI writing, killing the "this isn't X, it's Y" flip, stripping em dashes and hedging stacks, replacing generic claims with verifiable specifics, and matching an author's real voice.4---56# De-AI your content78Applies to every AI-assisted draft before publishing or sending. Produces the same copy with the machine tells removed, the vague claims made concrete, and a full detection battery passed before it ships.910The economics: readers pattern-match AI tells in about a second and discount everything after, feeds suppress the formulaic shape, and answer engines never cite the median take. An unedited AI draft is the one piece of content guaranteed to look like everyone else's. This pass injects the only thing that differentiates, which is structure a machine would not default to and specifics a machine could not know.1112## The pass, in order1314The order is the method. Reversing it produces a draft that has been word-swapped and still reads as generated.15161. **X-ray the skeleton.** Structural tells expose a draft even after every suspicious word is swapped, so structure gets fixed first. Scan against category 1 of the inventory and rewrite the sentence, never just the word.17182. **Strip the surface.** Run categories 2 through 5: punctuation, lexicon, content tells, voice tells.19203. **Run the specificity pass.** Every generic claim becomes a real number, name, or date pulled from source material, or it gets cut. If the source lacks the specific, flag the gap. Fabricated specificity is worse slop than vagueness, and in regulated industries it is a liability.21224. **Run the rhythm pass.** Read the piece aloud. Vary sentence length on purpose. Leave one fragment. Break at least one parallel structure per section. The target is copy that reads like one person typed it in one sitting.23245. **Reconcile with the author.** If a voice fingerprint document exists, load it and let it win every conflict with the tell list. Some humans genuinely open with rhetorical questions. Sanding off real habits produces a second kind of fake their audience catches just as fast.25266. **Calibrate for the channel.** Feed posts, cold email, long-form, and proof content each punish different tells.27287. **Prove it clean.** Run the full detection battery. A draft ships when it passes every applicable test, and never before — the correlation check is the one exception, since it only applies to teams running several accounts.2930## The reference pages3132- `references/tell-inventory.md` carries the full catalog in five categories, ordered by damage, with the fix for every entry. Structural findings mean rewriting sentences and merging paragraphs, and that happens before any surface fix, or the surface work gets overwritten.33- `references/channel-calibration.md` covers feed posts, cold email, newsletters and long-form, and case studies and proof content.34- `references/detection-tests.md` carries the seven tests, what each one catches, and what a repeated failure means.35- `references/voice-fingerprint.md` covers building the author document that overrides the tell list.36- `references/worked-examples.md` walks three passes end to end, plus one case where a flagged construction was correct and stayed.3738## The rule underneath the lexicon3940Any phrase that appears in everyone's AI drafts belongs to no one's voice. That is the test to apply when a word is not on the list yet.4142The same logic governs the specificity pass. A claim only a specific operator could make is the whole asset. A claim any competitor could publish unchanged argues for the category and says nothing about the company.4344## What good looks like4546The best operator reads for skeleton first: flips, triads, uniform rhythm. The mediocre version swaps "delve," deletes the em dashes, and ships, leaving the skeleton intact. The skeleton is the tell.4748The output is good when a reader who knows the author picks the draft out of a lineup at no better than chance, every specific traces to source material, at least one line costs the author something, and the author's actual quirks survived the pass.4950When a draft keeps failing the lineup or the x-ray after repeated passes, the problem is upstream. The source material is too thin or the voice fingerprint is missing, and no amount of polish substitutes for either.5152## Rules5354- MUST rewrite structure before touching word choice.55- MUST trace every number and name to source material, or cut it.56- MUST run every applicable detection test before anything ships (the correlation check applies only to multi-account teams).57- NEVER compensate with slang, forced edge, or manufactured casualness.58- NEVER treat paraphrase alone as humanizing. It preserves the structure that gives the draft away.59- NEVER override an author's real voice with the tell list.60- NEVER invent a specific to fill a gap. Flag the gap.
swan-gtm