The five points where the rules apply
| Where | What happens |
|---|---|
| In the prompt | The writing rules close the system prompt of every AI task whose output people read: the product profile, personas, narrative, messages, creative variants, account briefs and blog articles. They are a concrete list of banned words and sentence shapes, because models follow lists better than adjectives |
| After generation | Every string the model returns is cleaned. Tells that remain get one retry in the same conversation, listing each string and the rule it breaks. If the retry still reads as generated, the deterministic template answers instead and the call is logged with the reason |
| In the lint | Every creative’s copy is checked, including the text drawn into images, slides, carousel cards and video scenes. The automatic fix starts with the clean-up described below |
| In the editor | Each field shows what the detector finds as you type. On save the mechanical fixes are applied and listed, for example “3 dashes and 1 inflated word fixed automatically”, and anything that needs a rewrite stays an issue on the creative |
| Before sending | E-mail steps, organic posts, ads, bot replies and assisted DM drafts are checked again as they leave. Whatever still reads as generated does not go out |
What gets flagged
The detector looks for the patterns that give generated text away. They are described here by kind, since this page passes the same check.
| Tell | What it covers | What happens |
|---|---|---|
| Dashes | Em and en dashes used as punctuation, a spaced hyphen between words, arrows used as connectors, dashes used as bullets | Replaced with a comma, full stop or colon; a dash inside a number range becomes a hyphen |
| Inflated words | A fixed list of words with a plain equivalent, such as the long synonyms of use, help, simplify and important | Replaced with the plain word |
| Filler | Throat-clearing phrases and adverbs that only add emphasis | Removed |
| Stock phrases | Hype words, verbs that promise a transformation, stock e-mail openers and follow-up phrases, essay closers | Held for a rewrite |
| Stock constructions | The slogan that knocks down a straw man first, a question answered by its own next word, the colon reveal, scene-setting openers, three one-word sentences, rhyming triplets, a closing question about being ready | Held for a rewrite |
| Engagement bait | One-word questions that ask for agreement, and lines that ask readers to reply in the comments or to name a friend there | Held for a rewrite |
| Tell-tale emoji | A fixed set such as rockets, sparkles, fire, pointing hands and check marks | Removed |
Links, e-mail addresses and personalisation tokens are never matched or rewritten, so a stock word inside your own URL survives. Phrase and construction matching is English only; the dash rules apply in every language. The guide to AI tells in copy goes through the list with examples.
Deslop: only the fixes that keep the meaning
The clean-up step (called deslop in the code) only makes changes that cannot alter what a sentence says. A dash becomes a comma or a full stop depending on how long the clause after it is, an inflated word becomes the plain one, filler and those emoji go, and sentence capitals are tidied afterwards. It gives the same result every time, and running it twice changes nothing.
Everything else is a blocking issue with the offending words highlighted. A contrast slogan or a stock hook can only be fixed by writing a different sentence, so the model gets one try at that and a person gets the editor.
What happens at the last step
-
01
E-mail sequences
A step that still reads as generated is not sent. The enrolment is held with the reason and one review task is created per sequence step, not per recipient; other steps carry on. Editing the e-mail updates every step built from it and resumes the held enrolments.
-
02
Organic posts
Text, title and alt text are cleaned at publish time. A post that still fails is marked failed with the reason in the Social tab.
-
03
Paid ads
Ads go out cleaned. An ad whose copy still needs a rewrite blocks planning, the launch request and the launch itself.
-
04
Bot replies
The reply is not sent and the conversation is handed to a person, with the draft in the task.
-
05
Assisted DMs
Drafts are cleaned when they are created and again when they are shown. A draft with tells left cannot be approved for sending.
-
06
Stored creatives
When the rules change, approved and pending creatives are checked again once. Those the mechanical fixes clear are updated with a snapshot; the rest go back to approval with the issues listed.
The same check runs on this site
Every page of this website goes through the same detector before each deploy, and one hit on any page fails the check. The deterministic templates Demander uses when no AI is available are tested against it too, so template copy passes unchanged.
Questions people ask
Does it check copy my team writes in Demander?
Yes. Edits get the mechanical fixes on save and are checked again when they are sent or published, whoever wrote them. Replies people send from their own inbox are not touched.
Does it work for Czech or German copy?
The dash rules apply to every language. Phrase and construction matching is English only, so copy in other languages passes those checks untouched.
Can the rules be switched off?
There is no setting for that. If a flagged pattern is something you need, rewrite the sentence in the editor; the issue shows exactly which words triggered it.
What happens when the model keeps producing tells?
It gets one retry with the list of strings and rules. If the retry still fails, Demander uses its deterministic template for that task and records why, so you get usable copy either way.