Skip to content

Latest commit

 

History

History
55 lines (38 loc) · 3.32 KB

File metadata and controls

55 lines (38 loc) · 3.32 KB

PRD — Message Variant Engine

Problem

The moment a message has to reach more than one audience, brand voice starts to fragment. A marketer rewrites it for enterprise, a CS lead softens it for churned users, a regional team localizes it for APAC — and three weeks later the brand sounds like five different companies. The rewrite work is invisible, unowned, and impossible to keep consistent by policy alone.

Solution

Message Variant Engine encodes your brand voice as a reusable AI skill, then generates persona-, region-, and channel-specific variants from a single master message. Sections you mark as locked (legal language, product names, CTAs, claims) stay verbatim; everything else flexes to the audience. The brand owner defines the system once; anyone on the team produces on-brand variants in seconds.

Users

Brand & design owners (who define and protect the voice) and non-designers — marketers, customer success, sales, regional teams — who need audience-specific copy without going off-brand.

Primary personas

  1. The Brand/Design Owner — defines the voice, rules, and standards. Wants leverage: to encode their judgment once and have it applied everywhere without becoming a bottleneck.
  2. The Non-Designer Producer — needs on-brand output but lacks design/brand training. Wants speed and confidence that what they ship won't be "off."

Goals

  • G1 — On-brand by default. Output should be consistent with the brand system without the user having to know the rules.
  • G2 — Self-serve. A non-designer completes the task end-to-end without escalating to the design team.
  • G3 — Owned system. The brand rules live as a readable, versionable reference framework (markdown), not tribal knowledge.
  • G4 — Zero backend. No accounts, no server, no data collection — trust and auditability by design.

Requirements

Functional

  • Accept the user's input: Master message, target personas, target regions/locales, target channels, and any locked (verbatim) sections.
  • Produce: One on-brand variant per persona × region × channel combination, with locked sections preserved exactly.
  • Let the brand owner configure voice/rules that constrain every generation.
  • Store settings locally; never require an account.

Non-functional

  • Privacy: all data stays on-device except the direct, user-keyed call to Anthropic Claude (bring-your-own API key).
  • Latency: a full generation completes in a few seconds.
  • Auditability: the entire behavior is open source and inspectable.
  • Extensibility: the rule set is modular markdown, editable without touching code.

Success & adoption metrics

Metric Definition Target signal
Activation User completes one real generation First-session success
Self-serve rate Tasks completed without design-team review ↑ over time
On-brand acceptance Output shipped with minimal edits Low edit distance
Repeat use Return within 7 days Habitual use
Rule reuse Brand config reused across generations Encoded voice pays off

Non-goals

  • Not an autonomous agent, and not an evals harness. This is an AI skill: a structured, human-directed instruction set that produces reliable, on-brand output.
  • Not a content database or CMS.
  • Not a replacement for the brand/design owner — a force multiplier for them.