Verve
Strips AI tells from prose and puts a voice back
In British English, and it does the harder half. Subtraction alone leaves text that still reads as machine-made, just blandly rather than floridly. Meaning is a veto: every figure, name and claim survives. Writing that already reads as human comes back untouched.
Install Verve
Add the DBHQ marketplace, once
/plugin marketplace add dbhq-uk/marketplaceInstall the skill
/plugin install verve@dbhq
Both commands work in Claude Code and in Codex. The marketplace is a public repository you can read first: dbhq-uk/marketplace (external site)
What it is good at
Doing both halves, not just the subtraction
Strip the tells and stop, and you get clean prose that still reads as machine-made - blandly rather than floridly, because nothing is behind it. So the pattern sweep is followed by a voice pass: opinions, rhythm variance, specificity, acknowledged complexity, a bit of mess.
Treating meaning as a veto
Every fact, number, name, date and citation survives unchanged, technical terms keep their exact wording, and nothing is invented. The result is scored on six dimensions and fidelity below 9 forces a revision whatever the total says, because that failure has changed what the text means rather than how it reads.
Knowing who the writing is for
Explaining a term to someone who uses it daily is rude, because it says you did not think they would know. Leaving it unexplained for a newcomer is unhelpful. Same sentence, different reader. Give it an audience in plain words and it reads off how much explanation survives, how far it may direct rather than offer, and how long the thing should be.
A politeness check that runs both ways
Ten patterns are judged against that reader on every pass, from glossing a term they use daily to instructions aimed at somebody senior. The other direction is checked too, because a guard built only against condescension turns everything curt - so the greeting, the thanks and the one apology that is owed all stay.
British English that stays British
That is the default and it holds throughout. Ask for American and you get American; hand it a draft that is already American and it keeps it that way. Code, quotations, proper nouns, titles and standards text are never converted whichever variety you choose.
Leaving good writing alone
Triage comes first, and returning your text unchanged is a valid result - with one line saying so. A tool that processes clean prose for the sake of processing it makes writing worse.
What it does not do
It is not a detector-evasion tool
It does not tune text to score against an AI classifier, and it will say so if asked. Earlier versions shipped an optional commercial detector-evasion engine; it was removed in July 2026 rather than kept as a switch nobody had to flip. It contradicted the free, read-every-line argument the rest of the skill is built on, and it aimed at the wrong target. A classifier score is not the goal; prose a person would be happy to put their name to is. Those two come apart the moment you optimise for the score, and when they do the score wins and the writing loses.
It will not change what the text says
Not a fact, not a claim, not a step in the argument, at any strength setting - including heavy, which rewrites most sentences from scratch. Aggressive rewriting tempts a model to compress three real points into one punchy line, and that is the failure the fidelity veto exists to catch.
It will not quietly convert your spelling
Silently turning somebody's color into colour is exactly the kind of unrequested edit verve exists to avoid, so a draft that is already American stays American unless you ask otherwise.
It rewrites one piece at a time
There is no whole-repository rewrite. A separate read-only audit scores which files are worth a run and hands back a table, worst first, plus the files it thinks you should leave alone. It changes nothing; you decide what gets a pass.
Read it before you install it
Every line is public and MIT licensed. The README is the long version of this page, including the parts that are still rough.