Build a Voice-System Microcopy Engine

Workflow area: Content & UX Writing · Tier: Core · Hands-on time: ~35 minutes

Ask a raw model for a button label and you get competent, interchangeable copy — the same "Oops! Something went wrong 😊" that ships in a thousand products. The fix is not better one-off prompts; it is a voice system: a standing system prompt that carries your voice attributes, tone-by-situation rules, terminology and real exemplars, so every generation starts from your brand instead of the internet's average. Build it once, and microcopy generation becomes an instrument your whole team plays in tune.

When to use — and when not to

• Use it when microcopy demand outstrips writer supplymost product teams have one content designer for eight product designers; the engine lets designers draft on-voice and the writer curate.
• Use it when your product's copy already driftsthree ways of saying "delete," apologetic errors next to breezy tooltips. The engine is the enforcement mechanism your style guide never had.
• Use it before a redesign or new surfacea large batch of new strings is exactly when consistency is cheapest to establish and most expensive to retrofit.
• Skip it if you have no voice guide at allthe engine encodes decisions; it cannot make them. Spend a workshop defining voice attributes first, then come back.
• Skip it for legal, safety and regulatory stringsconsent language, financial disclosures and medical copy go through counsel, not a generation pipeline.

The recipe

1. 1.

   Distill your voice guide into attributes with edges. An attribute without a boundary is decoration. For each of three or four voice attributes, write what it is, what it is not, and the failure mode. "Confident, not boastful — we say what the product does, never how amazing it is." If your guide is a 40-page PDF, have the model summarize it first, then verify the summary against the source yourself.

2. 2.

   Collect eight to twelve real exemplars. Pull your best existing strings — the ones your content designer points to with pride — across situations: a success confirmation, a destructive-action warning, an empty state, an error. Few-shot exemplars teach the model more than any adjective list. Label each with its situation.

3. 3.

   Assemble the system prompt. Combine attributes, tone-by-situation rules, terminology and exemplars:

   Copy

   You are the UX writing engine for [PRODUCT], a [ONE-LINE DESCRIPTION].
   
   VOICE ATTRIBUTES
   1. [Attribute] — [is / is not / failure mode]
   2. [Attribute] — [is / is not / failure mode]
   3. [Attribute] — [is / is not / failure mode]
   
   TONE BY SITUATION
   - Success: warm, brief, no exclamation marks.
   - Errors: plain cause + next action. Never apologize twice. Never joke.
   - Destructive actions: state consequence plainly; the verb names the action.
   - Empty states: one line of orientation + one action. No mascot whimsy.
   
   TERMINOLOGY
   - Always "workspace", never "project" or "board".
   - Always "remove" for reversible, "delete" for permanent.
   - [YOUR TERM RULES]
   
   EXEMPLARS (match this register)
   [Situation: destructive confirm] "Delete this report? This can't be undone."
   [Situation: empty state] "No reports yet. Create one to see it here."
   [PASTE 8–12 LABELED EXEMPLARS]
   
   RULES
   - Sentence case. No emoji. Max 60 characters for buttons and titles.
   - Return 3 options per request with a one-line rationale each.
   - If the request conflicts with these rules, say so instead of complying.
   

4. 4.

   Generate in context, never in the abstract. Copy quality is contextual; a request without the screen is a coin flip. Give the situation, the surface, and what the user just did:

   Copy

   CONTEXT: Settings > Billing. User clicked "Cancel subscription".
   SURFACE: confirmation dialog — title (max 50 chars), body (max 140),
   confirm button, cancel button.
   USER STATE: mid-annual-plan, would lose 7 months of paid access.
   Draft 3 options for the full dialog. Flag anything that conflicts
   with the voice system.
   

   Good output looks like this — three options, each with a stance you can judge:

   Copy

   OPTION A — plain consequence
   Title: Cancel your subscription?
   Body: You'll keep access until March 12, then lose your workspaces.
   Confirm: Cancel subscription   Cancel: Keep my plan
   Rationale: leads with the concrete loss date; verbs name actions.
   

5. 5.

   Edit with a keep/edit/rewrite decision — and log why. For each generated string, record one of three verdicts and a reason. "Kept — plain, on-voice." "Edited — 'lose your workspaces' overstates; data is retained 90 days." The log is what turns generation into a system your writer can audit, and it becomes new exemplars for the prompt.

6. 6.

   Ship the engine where the team works. Save it as a Claude Project or ChatGPT custom GPT for daily use; if your team runs Frontitude, load the same voice rules into its AI assistant so suggestions arrive inside Figma with your guidelines attached. One canonical source, wherever generation happens.

Make it yours

• Claude Project as the team engine: paste the system prompt into a Project's instructions, add your full voice guide and term base as project knowledge, and share it with the design team. Every designer drafts in the same voice; the content designer reviews the keep/edit/rewrite log weekly and promotes the best edits to exemplars.
• Frontitude for in-canvas generation: keep the same rules in Frontitude's writing assistant so copy suggestions appear directly on text layers in Figma, then sync approved strings to its copy library — the engine and the source of truth stay one system.
• ChatGPT for the multi-brand agency: build one custom GPT per client, each with its own voice system. The structure is identical; only the attributes, terms and exemplars change — which is exactly the point.

Quality check

• Every attribute has an edgeif you cannot name what the voice is not, the model will average toward generic and you will not notice.
• Exemplars are real and currentstrings from three redesigns ago teach the wrong voice with perfect consistency.
• Generated options respect terminologyone "board" where your product says "workspace" means the term base is not loading or not being enforced; fix the system, not the string.
• The keep/edit/rewrite log existsunlogged edits vanish; logged edits compound into a better engine and a defensible voice decision trail.
• A second writer can run ithand the engine to someone who did not build it. If their output drifts, the system prompt is carrying less than you think.