The problem with brand guides
Most brand voice documents are written once and never tested. Someone writes "be conversational, avoid jargon" and calls it done. Nobody checks whether an AI model can follow it, because nobody built a loop to check.
So I built one. It's a Claude Code skill called /unslopify.
What it does
/unslopify takes any piece of text and rewrites it against a brand guide, a plain .md file with rules like "no em dashes" or "no self-announcing as a thought leader." Then I give it feedback, which means I rewrite whatever it produced. The skill reads my rewrite, infers what rule I was applying, and amends the guide. Then we do it again.
loop:
draft = unslopify(text, brand_guide.md)
edit = human_rewrite(draft)
brand_guide.md = infer_and_merge_rules(edit)
if human.says("I'm happy now"): break
Three full passes on three different AI-generated posts got it to where it produces slop I have no objections to. That's a lower bar than good writing, and it's still the floor most brand voice tools never reach.
Why this is the skill
Once the guide holds enough rules drawn from edits I made by hand, I can hand Claude Code a bare instruction like "write a short LinkedIn post about building Unslopify" and it produces something close on the first try. If it misses, I run /unslopify on the output instead of rewriting from scratch. The correction loop doubles as the editing tool.
Build time: about 10 minutes with Claude Code, Opus, high effort. The thing that takes more than 10 minutes is noticing what you object to, sentence by sentence, every time.
The brand guide is a record of every fight I've had with a draft.
The guide only works because the rules in it are testable, which is the same argument as The Kindness of Definition. And the loop itself is a review loop, so Review Is a Contract covers why the correction step is the part carrying the weight.