Step 1

Replace adjectives with contrasts

Define voice through concrete choices: concise rather than breathless, informed rather than technical, confident rather than absolute, playful without sarcasm.

Pair each principle with approved and rejected examples drawn from real product language.

Words such as authentic, premium, and friendly are too broad to produce consistent decisions without examples.

Working checklist

  • Write four voice contrasts
  • Add approved examples
  • Add rejected examples
  • Define punctuation rules
Step 2

Create a claims boundary

List which product facts may be restated, which require a source, and which the system may never infer. Include health, sustainability, scarcity, guarantee, and comparative claims.

Make missing data produce an omission or a review flag rather than plausible-sounding filler.

Generative confidence is not evidence. A fluent claim can still be false or legally sensitive.

Working checklist

  • Classify claim types
  • Link approved sources
  • Define missing-data behavior
  • Escalate sensitive categories
Step 3

Build reusable message patterns

Create patterns for launches, features, care tips, comparisons, restocks, and social proof. Each pattern should preserve room for product-specific evidence.

Vary hooks and structures within a pattern so consistency does not become detectable repetition.

Do not use the same opening, emoji sequence, and CTA across the entire catalog.

Working checklist

  • Create six message patterns
  • Provide variation rules
  • Set CTA options
  • Limit repeated phrases
Step 4

Audit live output

Sample posts by product category and template. Compare them with the source data and brand examples, then record drift.

Refresh the guide when the assortment, positioning, audience, or channel norms change.

A static prompt slowly becomes inaccurate even if the model itself does not change.

Working checklist

  • Sample ten posts monthly
  • Track drift
  • Update examples
  • Retire weak patterns