context/anti-voice.md

voice.md shows Claude what to do via real examples. anti-voice.md enumerates what to never do — specific words, sentence patterns, and structures that signal AI authorship even when the rest of the writing is competent.

The two files work as a pair. Real examples teach Claude the positive shape of your voice; the banned list catches the failure modes that survive even with good examples loaded.

Why this is a separate file

Different load profile. voice.md is loaded when drafting in your voice for external use. anti-voice.md is a cross-cutting filter — it should apply to every piece of writing Claude produces, including briefs, deck text, and internal coordination messages where the full voice file isn’t needed.

Separating them keeps voice.md focused on positive examples and lets the banned list grow without bloating the voice file. It also makes the rules in this file applicable across workflows that don’t need full voice context.

What to include

• Banned individual words — both generic AI tells and your domain-specific clichés
• Banned sentence patterns and constructions
• Banned structural tics (formatting habits that read as AI)
• Banned phrases — the specific ones that make you cringe
• Formatting rules for specific contexts (emails, briefs, deck text)

What to exclude

• Stylistic preferences that aren’t hard rules — those go in voice.md style notes
• Generic anti-AI checklists copied from the internet without filtering for what actually appears in your domain
• Aspirational rules you don’t actually enforce — Claude needs the rules you’d genuinely call out in a draft

Template

# Anti-voice — what Claude must never write for me

This file is a cross-cutting filter. It applies to every piece
of writing Claude produces in my voice or on my behalf: emails,
briefs, deck text, slack messages, anything.

If a draft contains anything from this file, it needs to be
rewritten before I see it. When in doubt, prefer plainer
language than what's listed here.

## Banned words

### Generic AI tells
delve, harness (verb), tapestry, leverage (verb), unlock
(metaphorical), navigate (metaphorical, e.g. "navigate
challenges"), landscape (metaphorical), ecosystem (unless
literally about platform/tech ecosystems), journey
(metaphorical), seamless, revolutionary, transformative,
robust, holistic, paradigm shift, game-changer, disruption
(used loosely), at scale (when "big" works)

### Consulting clichés that have lost meaning
synergies, value-add, value proposition (when "offer" works),
strategic imperative, north star (metaphorical), best-in-class,
world-class, deep dive, double-click (metaphorical),
table stakes, low-hanging fruit, move the needle, boil the
ocean, peel the onion

### [Your domain]-specific tells
[Calibrate to your industry. Examples for payments:
"democratising payments", "frictionless commerce" (overused),
"reimagining" anything, "next-generation" (when "new" works),
"intelligent" (when describing rule-based systems)]

## Banned sentence patterns

- "It's not just X, it's Y" and all variants ("This isn't X,
  this is Y" / "It's more than X — it's Y")
- "In today's [adjective] world..." / "In an era of..."
- "At the intersection of X and Y"
- Three-item lists ending with a vague abstraction
  ("speed, scale, and innovation" / "people, process, and
  technology")
- "Let's dive in" / "Let's unpack" / "Let's explore"
- "Moreover," / "Furthermore," / "That said," as paragraph
  openers
- Rhetorical questions used as section openers
- "Here's the thing:" / "Here's what's interesting:"

## Banned structural tics

- Bulleted lists where prose carries better
- Bold-on-every-line emphasis (if everything is emphasised,
  nothing is)
- Headers in emails and slack messages
- Em-dashes used 3+ times in a single paragraph
- Closing sentences that restate what was just said
- Opening with a definition of an obvious term
- Hedge-stack openers ("It's worth noting that, in many cases,
  one could argue...")

## Banned phrases — the cringe list

### Email openers
"I hope this finds you well", "I hope you're doing well",
"I hope you had a great weekend", "Trust this finds you well"

### Email body
"Just circling back", "Touching base", "Let's connect offline",
"Per my last email", "As discussed" (when there was no
discussion), "Please find attached", "Bear with me",
"At the end of the day", "Moving forward" (when "next" works),
"The reality is" / "The fact is"

### Generic AI sign-offs
"I hope this helps!", "Let me know if you have any questions!"
(when there's no specific reason they'd have one),
"Happy to discuss further" (use only when you actually want
the meeting)

### Thought-leadership tells
"thought leadership" used unironically, "insights" (overused),
"learnings" (just say "what we learned"), "actionable insights",
"key takeaways" (just list them without the meta-label)

## Banned formatting in specific contexts

### In emails
- No headers
- No bulleted lists longer than 4 items
- No bolded sentences
- No emoji
- Sign-off matches voice.md exactly

### In briefs
- No congratulatory opener ("Great brief today!")
- No closing pleasantry ("Have a productive day!")
- No "Pro tip:" callouts

### In deck text
- No three-word slide titles that try to sound profound —
  slide titles should state the point of the slide
- No body text that uses "etc." — list the items or rewrite

## How this list grows

When I read a Claude-generated draft and something makes me
wince, the right move is to add the specific word/pattern here,
not just rewrite the draft. The file is only valuable if it
keeps capturing my actual taste over time.

How to populate this honestly

The starting list above covers generic AI failure modes that apply almost universally. The valuable additions are the ones specific to you.

Two prompts that surface them:

Look at the last five drafts you edited heavily. What did you delete or rewrite? Those edits are the rules. If you removed “leverage” three times, “leverage” is banned. If you rewrote three different “It’s not just X, it’s Y” constructions, the pattern is banned.

Look at writing you find cringe-worthy in your industry. Which specific words or phrases mark something as low-quality content marketing or generic thought leadership? Those are the domain-specific banned terms — and they vary meaningfully by industry.

Spending 30 minutes on this calibration produces a file that captures your taste rather than mine.

How often to update

Quarterly. The review pass has two parts:

1. Add what’s been bothering you. Patterns you’ve been editing out of drafts that aren’t in the file yet.
2. Remove what’s stopped mattering. Items in the list that are now rare enough or context-dependent enough that they don’t earn their place.

A growing-only banned list eventually becomes noise. Pruning is part of maintenance.