Your AI Can't Spell Your Company Name

Editor’s note (March 2026): Since this post, Coherence has shipped v1.6 with a unified /coherence command and 11 sub-commands. The standalone /check-consistency and /check-drift commands described here are now /coherence check-drift and /coherence check-principles. /check-consistency has been retired — terminology and style enforcement now runs entirely through automatic hooks. See coherence.viewyonder.com for the current command reference.

Every GenAI tool writes confidently. Most of them write wrong.

Not factually wrong — that’s a known problem with known mitigations. I mean stylistically wrong. Wrong product names. Wrong tone. Wrong hyphenation. The kind of errors that make your brand look like it was assembled by a committee of autocomplete engines.

We ship a lot of content through Claude Code at Viewyonder. Articles, case studies, build logs, service pages. And we kept hitting the same problems: “View Yonder” instead of “Viewyonder”. “Game-changing AI solutions” instead of anything a human would say. Headings that jumped from H1 to H3 like the document had a structural death wish.

So we built guardrails. Not the “please don’t say bad words” kind. The kind that enforce your actual style guide, in real time, as the AI writes.

Two Layers: Automatic and On-Demand

The system has two modes. The automatic hooks run silently every time Claude edits a file. The manual commands run when you ask for them.

What Happens Automatically

Three hooks fire on every edit:

Terminology check catches product names and brand terms. Write “gen ai” and it gets flagged — the correct form is “GenAI”. Write “Injection-ator” and it suggests “Injectionator”. This covers about 15 rules: product names, AI terminology, hyphenation conventions.

Style guard flags marketing fluff. Words like “revolutionary”, “synergy”, “leverage”, “unprecedented” — the vocabulary of people who haven’t decided what they actually want to say. It also catches AI hype patterns: “AI will replace” gets flagged because our voice frames AI as augmenting people, not replacing them.

Change suggest watches for edits to critical files and reminds Claude to check related content. Edit a service page? It nudges: “Check if the services hub and homepage cards need updating.” Change the design tokens? “Check components that use the modified variables.”

The user experience is invisible. A marketing person asks Claude to write a blog post. Claude writes it. The hooks quietly correct terminology and steer the voice. The person sees brief status messages — “Checking terminology…” — and the output just comes out right.

What You Run Manually

Two commands, typed directly into Claude Code:

/check-consistency reviews content against the full style guide. Run it against all articles, a specific directory, or just your staged git changes. It checks terminology, voice, structure, and cross-references, then reports pass/warn/fail per category.

/check-consistency                      # Everything
/check-consistency src/data/articles/   # Just articles
/check-consistency staged               # Just what you're about to commit

/check-drift compares your SPEC documents against the actual codebase. It answers the question: “Has reality drifted from what we documented?” You can scope it to content standards or site architecture.

/check-drift              # Full check
/check-drift content      # Content standards only
/check-drift architecture # Site architecture only

The Stack

This runs on Claude Code’s hook system — shell scripts that fire on specific tool events. The hooks are Node.js files that parse the edit content, match it against pattern rules, and return warnings. No external services. No API calls. Everything runs locally.

The configuration is straightforward:

• PreToolUse hooks fire before an edit is applied — they can warn or block
• PostToolUse hooks fire after — they suggest follow-up actions
• Agents define review personas with specific expertise
• Skills are slash commands that invoke those agents

Currently our hooks warn rather than block. Claude sees the warning and self-corrects. For strict brand compliance, you could switch them to blocking mode — Claude would have to fix the issue before the edit goes through.

What a Content Person Needs to Know

Here’s the onboarding conversation:

1. Open terminal in the website folder, run claude
2. Ask Claude to write or edit content normally — it auto-corrects terminology and tone
3. Before publishing, type /check-consistency to review everything
4. If Claude flags a term, trust it — it’s enforcing the style guide

That’s it. Four bullet points. The guardrails handle the rest.

What’s Next

This is now Coherence — a shipped product at v1.6 with a unified /coherence command and 11 sub-commands that generate and manage a complete guardrails system for any project type. Web apps get architectural boundary enforcement. APIs get data isolation checks. Content projects like this one get terminology and voice guards.

The principle is simple: encode your standards once, enforce them on every edit. Your AI should know how to spell your company name.