Claude for Business: The Skill Stack That Ends AI Slop
I hate to break it to you, but one brand voice skill was never going to do the job.
You created a brand voice skill, installed it, fed it your tone words, your banned list, your entire manifesto about how you'd rather eat drywall than type the word "unlock." And Claude still slid an em dash into the closing line of a draft that was going so well you didn't bother proofreading to the end, and posted it anyway.
If you have ever screamed I SAID NO EM DASHES into the void, hi. This one is for you specifically.
I run the whole backend of my business through Claude, but my voice is only as good as my enforcement - not one drop better.
Back when enforcement lived in one skill, it slipped every third draft. Bulleted lists nobody asked for, sign-offs that went full customer service rep at the exact moment the piece started getting good, like the model got nervous and put on a blazer and some reading glasses.
Splitting the job into layers is what fixed it.
Here's the five-layer stack I ruse religiously, what each layer catches that the others sleep right through, and the edit that makes the whole system fire when it actually should.
TL;DR: voice enforcement is four separate jobs. What you sound like, the mechanics under the sound, the compliance check before anything is finished, and the format each platform needs. One skill trying to do all four will drop the ball on at least one every single time. That’s why I have five layers, each scoped to one job.
Check it out!
Why Isn't One Brand Voice Skill Enough?
A brand voice skill covers tone and vocabulary. Voice enforcement also needs rhythm mechanics, a delivery gate, and format moves that no single skill can manage on its own.
The failure is actually structural, and it starts with how skills are read.
A voice skill grpws every time you catch a new pattern. - and the longer it gets, the more Claude treats the whole thing like vibes instead of enforceable rules. The memorable ones that you mention a lot stick, the mechanical ones often get skipped, and then your voice starts to drift and you end up either arguing, or just redoing the whole damn thing yourself.
If your flavor of ADHD is the one that makes spot-editing physically painful, good luck to you and to me both. The likely result is that you end up scrapping the whole thing, then beating yourself up about it for weeks on end. (Hello, newsletters, I’m talkin’ ‘bout you!).
The jobs are also wildly different kinds of work.
Vocabulary is a lookup. Your voice rhythm has to make it through a whole draft without sagging, and verbiage compliance is a literal count that either hits zero or it doesn't - there’s no inbetween.
If you give one skill all of this it will handle the most obvious part, then skip the rest and give you shit output, like a roommate who "did the dishes”, then you come home to find a sink full of them. (You should probably get a new roommate if that’s the case, btw!).
A skill written to describe your voice has one job: to shape what Claude generates.
But nothing in that skill actually audits what came out. Which is the whole reason you're out here yelling at a language model in your home office while your pets look at you like the lunatic you are.
What Is the Voice Enforcement Stack?
The voice enforcement stack is five skill layers: brand voice, editing patterns, a final scrub gate, format-specific skills, and pair-loading.
Top to bottom:
Brand voice skill. The what of your voice. Always loads on writing tasks.
Editing patterns skill. The mechanics under the voice. Always loads, paired with brand voice.
Final scrub skill. The delivery gate. Runs last, right before anything is delivered.
Format-specific skills. Carousel, Threads chain, email, blog. Load only when the format matches.
Pair-loading. Skill descriptions that name their partners so the layers load together.
Every layer exists because the one above it has a blind spot.
What Does Each Layer Actually Do?
Layers one and two shape generation, layer three audits output, layer four adds format rules, and layer five wires the whole thing to load together as one system.
Layer 1: The Brand Voice Skill
The what. Tone and vocabulary, plus the patterns that make the writing recognizable as yours. I've got fifteen named voice moves in mine, each one with annotated examples pulled from real posts I’ve written, the actual audience names I use, and a list of coach-speak I wouldn't touch with somebody else's mouth. It loads on every writing task. No exceptions or vibes-based judgment calls, it shows up every time.
Its limit: it describes what your voice sounds like, and description alone will not stop the model's defaults from creeping back in around the edges. Which usually happens at the final line, and right after your attention has officially left the building, so you either end up missing it or catching it and posting it anyway because exhaustion.
Layer 2: The Editing Patterns Skill
The mechanics. Rhythm rules, punctuation preferences, sentence-opener bans, capitalization defaults.
Vocabulary bans are the easy part, the rhythm is where AI writing turns itself in at the front desk.
Stacked staccato sentences, every paragraph built to the same shape, that one cadence that reads like a LinkedIn post written by James in accounting.
Mine bans three short sentences in a row and pins the em dash count at a flat zero, plus about a dozen other mechanical checks I've added one funeral at a time.
It loads alongside the brand voice skill every single time, because the sound and the mechanics under the sound are two different jobs that need to happen in parallel.
Layer 3: The Final Scrub Skill
The delivery gate. This one runs literal string matches against the ban list. Em dashes, banned AI vocabulary, sycophancy openers, every phrasing in my past-mistakes ledger, which is longer than I'd like and still growing.
It runs before the draft streams and rewrites silently until the count hits zero.
The first two layers are constraints on generation. This layer is an audit on output, and that matters because the two fail in completely different ways. A model can keep your voice fully intact for 800 words and still go full brand account in the sign-off. The scrub exists for that and that alone.
Layer 4: Format-Specific Skills
Instagram carousel, Threads chain, email copywriter, blog writer. Each one loads only when the format matches the request, and each stacks on top of the always-load pair, so the format skill fires right alongside the universal ones.
My blog skill enforces answer capsules and question-formatted headers. My Threads skill knows chain pacing, which is its own animal.
None of them re-teach, or even mention the voice. They inherit that from the layers underneath, which keeps every skill small and keeps Claude from getting two contradictory rules and picking a winner at random on its own.
Layer 5: Pair-Loading
The wiring, and the layer hardly anyone thinks of.
Claude reads skill descriptions when it decides what to load, so the description is where the stack gets assembled. Mine name their partners outright. "Always load alongside brand-voice and editing-patterns." "Load for carousel requests only, never for single-image posts."
Once the pairings live in the descriptions, all five layers load on every request without you invoking a damn thing. The stack stops depending on your memory, which, if your memory runs anything like mine, was never a load-bearing wall.
what keeps happening, and the missing layer:
Em dashes and banned words survive into final drafts.
Missing layer: the final scrub skill.
The words are right but the rhythm reads robotic.
Missing layer: the editing patterns skill.
Output sounds like a competent stranger wrote it.
Missing layer: the brand voice skill.
Voice holds in one format and collapses in another.
Missing layer: format-specific skills.
Everything works, but only when you remember to ask.
Missing layer: pair-loading.
Two of these get misdiagnosed seventy eleven times a day, so let's sit with them.
Output that sounds like a competent stranger is a description problem. The instinct is to shove more banned words onto the list, but that instinct is wrong. I know that because I did it for months and produced a ban list long enough to qualify as a novella.
Rhythm that reads robotic is a mechanics problem. The instinct there is to rewrite the voice description for the fifth time (also wrong). Both fixes fail because the fix lives in a different layer than the one getting edited, and no amount of editing the wrong layer will change that. Ask me how I know that one too.
How Do You Build the Stack on a Low-Capacity Day?
Start with the delivery gate. That one scrub skill will the worst slop offenders right away - you can add the rest of the layers one at a time as capacity allows.
You don’t need to build all five layers in a weekend, and if your capacity runs variable like mine does, please do not try. Your body will send an invoice your energy bank won’t be able to pay.
The low-capacity path. Install a scrub layer first. Then add layers one at a time in whatever order matches what pisses you off most. Editing patterns next if rhythm is your tell. A format skill next if one platform is running your whole business on its back.
The higher-capacity path. You can do it in one sitting, roughly a little over an hour. Pull up every skill in your setup, split anything monolithic into a voice skill and a mechanics skill, then rewrite every description with its load condition and its partners named. That last part is fifteen minutes doing most of the work, which is rude of it.
What Is the Fifteen-Minute Edit?
If you want the delivery gate without building it yourself, I built that layer for you and you can grab it for free!
Stoppy Sloppy is my free Claude skill that runs the scrub on every output,
catching the slop before it lands on the page. It slots into a full stack or runs alone next to whatever half-built setup you have today, and it's free:
