Why I Run Claude Code Plugins for Brand Voice Enforcement

Every article I publish runs through a plugin that reads my brand rules and rejects anything that breaks them. Last month it caught 31 em-dashes, 9 banned words, and 4 AI tells across 12 drafts. None of those reached the live site.

The Problem I Was Actually Solving

I write a lot. One person, one studio, dozens of articles a month across the lab. The hard part was never writing. It was staying consistent. By draft number eight my sentences would creep longer. An em-dash would slip in. I would write "use" instead of "use" because my brain was tired and the autocomplete in my head defaulted to the corporate version.

Brand voice is not a vibe. It is a list of rules. No em-dashes. Currency written as "33 EUR" not with a symbol. First person singular, never "we" when it is just me. No words that leak legal meaning. No tired AI phrases like "understand the space" or "today's world." I had all of this written down in a document. The document did nothing. A document is a suggestion. I needed something that would actually stop a bad draft from going out.

The first version was me reading every draft three times. That worked until it did not, because I am the writer and the reviewer and the publisher, and the reviewer gets tired at the same rate as the writer. I missed an em-dash in a published piece and only found it two weeks later. That was the moment I decided the rules needed to live in code, not in my attention span.

I tried doing this with a simple find-and-replace script first. It caught the obvious stuff but it had no context. It would flag "buffer" every time, even when I meant the scheduling product. It would not understand that "return" is fine in a sentence about return visitors but a problem in a sentence about returning a product. I needed something that could read the draft like a person and apply judgment, not just match strings. That pushed me toward a plugin running inside Claude Code, where the model itself does the reading and the plugin enforces the verdict.

Why a Plugin and Not a Skill

This is the distinction that took me a while to understand, so I will be plain about it. A skill is something Claude can do when you ask. A plugin is something that runs on a trigger whether you ask or not.

If brand voice were a skill, I would have to remember to invoke it. "Check this for banned words." But the whole point is that I forget. The tired reviewer forgets. A check I have to remember is a check I will skip on the day it matters most, which is the day I am rushing a piece out before a deadline. Skills are great for things you choose to do. Brand enforcement is not a choice. It is a gate.

So I built it as a plugin that hooks into the moment before a draft is considered done. The plugin reads the full text, runs it against my rule set, and returns a list of violations with line references. If the list is not empty, the draft does not move forward. No human discipline required. The discipline lives in the hook.

The rule set is not complicated to maintain. It is a structured list: banned exact phrases, banned characters (the em-dash being the famous one), currency format patterns, voice rules like first person singular, and a list of AI tells with plain-English replacements. When I notice a new bad habit in my own writing, I add a line. The plugin gets smarter every week without me rebuilding anything.

The part that makes the plugin worth more than a regex is context. When it sees "buffer," it checks whether I mean the scheduling tool, and if I do, it confirms I included the link. When it sees a long sentence, it does not just flag length, it suggests where to break it. That judgment is the model doing the reading, wrapped in a plugin that turns the reading into a pass-or-fail decision. If you want the full picture of how I wired Claude Code in the first place, I walk through it in the Claude Blueprint.

What It Catches That I Miss

Here is the honest accounting. In the last 12 drafts the plugin caught 31 em-dashes. I genuinely do not see them anymore when I type, because the keyboard shortcut is muscle memory and the character looks fine in the moment. The plugin sees all 31. Every single one.

It caught 9 banned words. Some of these are legal-leak terms I will not even type here because they create the wrong impression on a commerce site. Others are voice violations: "use" three times, "solid" twice, "smooth" once. These are not wrong words in the abstract. They are wrong for me. They are the words a brand sounds like when nobody is steering it, and my whole point is to sound like a person.

It caught 4 AI tells. "Note" appeared once. "Go into" appeared once. "Understand the space" appeared twice in the same draft, which tells you exactly how tired I was that day. The plugin does not just delete these. It rewrites them into plain language and shows me the swap, so I learn the pattern instead of leaning on the tool forever.

The currency check is small but it matters. I sell things. Prices show up. If I write a price with a symbol in one article and as "29 EUR" in another, the site looks careless. Careless on a price is the worst place to be careless. The plugin normalizes every price to the same format, so a reader sees one consistent style across everything.

The numbers that convinced me to keep it: zero brand-voice violations have reached a published article since the plugin went live. Before it, I was finding roughly one mistake per published piece on re-read, usually too late to matter. Going from one to zero sounds small. Across a year of articles it is the difference between a site that reads like a studio and a site that reads like a draft.

How It Fits the Rest of My Workflow

The plugin is one gate in a longer pipeline, and it matters where it sits. It runs after the writing is finished but before anything touches the live store. By then the article exists in full. The plugin reads it once, returns its verdict, and either passes the text along or sends it back with the violation list.

What I like is that the same rule set governs everything. Long-form articles, shorter posts, even the copy that goes out when I schedule social. When I push a piece to Buffer for scheduling, the social copy goes through the same voice check first, so a post never contradicts the tone of the article it links to. One source of truth for how I sound, applied everywhere, automatically.

The store side runs on Shopify, and the published article lands there with the formatting already correct, prices already normalized, voice already clean. There is no second pass where I fix things in the Shopify editor, because the fixing happened upstream in the plugin. That is the whole design philosophy: catch problems at the earliest point where they are cheap to fix, never at the latest point where they are embarrassing.

The plugin also keeps a log. Every violation it catches gets recorded with the date and the draft. Over a few months that log becomes a map of my own bad habits. I can see that em-dashes spike when I write late, that AI tells cluster in drafts I wrote fast, that banned legal terms creep in when I am writing about products rather than process. That data shapes how I write, not just what gets published. The tool stopped being a filter and became a mirror.

Maintenance is almost nothing. I add a rule line when I spot a new habit. I review the log monthly. The plugin does the rest on every draft, forever, without getting tired or distracted or rushed the way a human reviewer always eventually does.

Bottom Line

Brand voice is a set of rules, and rules that live in a document do nothing. The moment I moved my voice rules into a plugin that runs automatically before publish, my published mistakes went from about one per article to zero. The difference between a plugin and a skill is the difference between a gate and a suggestion. A skill waits for me to remember. A plugin runs whether I remember or not, which is exactly what you want from the thing standing between your tired late-night draft and your live store.

If you write at any volume and you are the only reviewer you have, build the gate. Put your banned words, your formatting rules, and your AI tells into something that reads every draft and refuses to pass the broken ones. The full setup, including how the hooks fit together, is in the Claude Blueprint. Start with five rules you break most. Add more as the log teaches you what you actually do wrong.

This article contains affiliate links. If you sign up through them, I may earn a small commission at no extra cost to you. (Ad)

Keep reading

AI ToolsOpus 4.8 vs Sonnet vs Haiku: How I Route Work in 2026 9 min read

AI ToolsOpus 4.8 in Practice: A Real Coding Session 9 min read

AI ToolsClaude Code Hooks I Ship in Every Project: 6 Patterns 11 min read