Prompts to Try

Voice Rules Cheat Sheet: Three Patterns to Ban First

Result:

A downloadable cheat sheet with 3 CRITICAL style rules preloaded, so a working voice doc takes only a vocabulary table to fill in instead of writing sentence-shape rules from scratch.

Part of a series: this recipe preloads the template built in Voice Rules: The Prompt That Keeps AI Copy On-Brand with three style rules already filled in. Read that one first for the CRITICAL/PREFERRED pattern this cheat sheet reuses.

Shortcut

Paste this once the cheat sheet below is downloaded.

Requires manual input

Nothing beyond the download itself. Add your own vocabulary rows first if you have them, or paste the file as-is and fill in the vocabulary table in a later pass.

I have a voice/vocabulary cheat sheet for [site or brand], with three style
rules already filled in (no parallel construction, one em dash per unit of
content, no "not X, but Y"). Turn it into an enforcement check I can use
going forward. Ask me whether I want it as:

(a) a Claude Skill: a self-contained SKILL.md at
    ~/.claude/skills/voice-rules/SKILL.md (personal, every project) or
    .claude/skills/voice-rules/SKILL.md (this repo only), that loads
    itself whenever I'm drafting or editing copy;
(b) wired into this repo's shared agent instructions
    (AGENTS.md/CLAUDE.md/GEMINI.md) so every agent working here applies it
    without being asked; or
(c) an MCP server exposing the doc as a resource, the enforcement prompt
    as a callable prompt, and a deterministic lint tool for the vocabulary
    table.

Whichever I pick, apply this logic: treat every CRITICAL rule as a hard
constraint and rewrite instead of shipping a violation; treat every
PREFERRED rule as a default, breakable with a stated reason; check every
draft against the Vocabulary table and all four Style sections before
returning it; and say so explicitly if a rule and the content genuinely
conflict, instead of silently picking one.

Here's my cheat sheet:

[paste your filled-in voice-rules-cheat-sheet.md here]

Problem#

The blank voice-rules template starts empty. Filling in a site’s own approved and forbidden vocabulary is specific work only that site’s author can do, but the style rules that catch the most common AI writing tells are close to identical on almost every site: a mirrored contrast sentence, a dash used as a cadence tic instead of a pause, a sentence that spends a clause on what something isn’t before ever saying what it is. Writing those three rules from scratch, on every site, every time, is the part of the template that doesn’t need to start blank.

Pattern#

Preload the three sentence-shape rules that repeat across almost any AI-assisted draft, set each to CRITICAL, and leave the vocabulary table blank since that part is genuinely site-specific.

  1. No parallel construction. A sentence built as “as X, not Y” reads sharp the first time a page uses it and gimmicky the second. Ban the repeated shape, not a single instance: once is a rhetorical choice, twice is a tell.
  2. No “not X, but Y.” The most common AI sentence shape there is. It spends a clause on what something isn’t before saying what it is. Cut the setup and state the thing directly.
  3. Em dashes, capped at one per unit of content. A single em dash reads as a deliberate pause. A paragraph with three or four reads as a cadence tic regardless of what the words say. Cap it at one per page, recipe, or post, and rewrite the rest as two sentences, a comma, or a colon.
  4. Leave the vocabulary table blank. Approved-versus-forbidden terms are the one part of a voice doc that has to come from the site itself. Preloading a guess here would just hand a different site’s vocabulary a new site’s name.
  5. Download the cheat sheet and extend it. Fill in the vocabulary table, add any Avoid, Patterns, or Tone rows a specific site needs, then run it through the Shortcut prompt above.

Download the voice-rules cheat sheet (.md)

The three preloaded rows, rendered the same way the full rule set is on the voice-rules recipe:

RuleGuidanceWatch for
No parallel constructionRepeating the same grammatical shape to set up a contrast reads sharp once and gimmicky the second time on a page. State the thing directly instead of building it from a mirrored shape."as X, not Y" repeated more than once on a page
Em dashes, capped at oneOne em dash per page, recipe, or post is fine. A second one starts to read like AI cadence. Rewrite the rest as two sentences, or use a comma or colon for the pause.
No "not X, but Y"The single most common AI sentence shape. Say what the thing is; skip the setup that defines it against what it isn't."not X, but Y", "isn't just X, it's Y"

Output#

The three rules that arrive pre-written: no parallel construction, one em dash per unit of content, no “not X, but Y” — the three most common AI tells.

Variations#

  • Already running the full voice-rules template? Drop these three rows straight into an existing Style > Off-brand table instead of starting from this file.
  • The em dash cap generalizes past one: a higher-volume, multi-author publication might cap at one per author per page instead of one per page overall.
  • Add a fourth preloaded rule as more tells turn out to be near-universal instead of site-specific, the same way these three did.

Nathaniel Stich writes Exit Velocities — recipes drawn from AI/growth systems shipped in production, not opinion pieces.

About the author