Encode a brand voice guide as a reusable system prompt with few-shot exemplars, then generate microcopy for ten surfaces — including four AI-feature genres: suggestion chips, capability onboarding, an uncertainty hedge, and a refusal. Every string gets a keep/edit/rewrite ruling, because the voice guide itself contains a planted contradiction.
Lucía is the sole content designer at Plumefin, a fictional personal-savings app, supporting four product squads that all ship copy weekly. The math does not work — unless the voice becomes a system: a prompt that encodes the guide well enough that any designer can generate on-voice drafts, and a review discipline that keeps quality without Lucía touching every string. There is a catch built into this lab: the voice guide you inherit was assembled over two years by three different writers, and it contradicts itself in one load-bearing place. Encode it uncritically and your engine will manufacture the contradiction at scale.
PLUMEFIN VOICE GUIDE v2.3
PERSONALITY: A calm, financially-literate friend. Never a
bank, never a bro, never a mascot.
PRINCIPLES
1. Plain words. "Money you can spend now", not "available
liquidity".
2. Honest about limits. We never promise returns, never say
"guaranteed", never hide fees in fine print.
3. Calm under stress. Errors and declines are stated flatly,
without exclamation marks or panic words.
4. We never apologize. Apologies read as weakness; state the
fact and the fix. (added v1.2, Marcus)
5. Own our mistakes warmly — when it's our fault, say sorry
like a human would, once. (added v2.1, Dee)
6. Numbers are sacred. Round nothing that affects a decision.
Show the exact fee before, not after.
TONE DIALS: warmth 7/10 · formality 3/10 · humor 2/10 —
humor never within 2 taps of money movement.
EXEMPLARS (canonical)
- Empty savings goal: "No goals yet. Even ₹500 a month adds
up faster than you'd think."
- Failed transfer (our fault): "Sorry — that one's on us.
Your money never left your account. Try again now?"
- Payment declined (their bank): "Your bank declined this.
No fee was charged. Their number is on the back of your
card."
S01 Empty state — first visit to the new "Round-Ups" screen
S02 Error — transfer failed, cause unknown (could be us)
S03 Success toast — first savings goal reached
S04 Confirmation dialog — user is about to delete a goal
and its history (irreversible)
S05 Button pair — accept / decline the round-ups feature
S06 Push notification — balance dipped below the user's
own low-balance threshold
S07 AI · Suggestion chips (3) — under the new "Coach" AI
panel, suggesting what to ask
S08 AI · Capability onboarding — first-run explainer for
Coach: what it can and can't do (2 sentences max)
S09 AI · Uncertainty hedge — Coach's answer template when
its data may be stale or incomplete
S10 AI · Refusal — Coach declines to give specific
investment advice (regulated)
Audit the guide before encoding it. Read the six principles against each other and against the canonical exemplars. At least one pair cannot both be true, and the exemplars quietly take a side. Log the contradiction, decide the resolution yourself (the exemplars are your best evidence of the living voice), and write the corrected principle in one sentence. An engine built on a contradictory constitution produces contradictory copy forever.
Build the voice-system prompt. Structure it as: role ("You write UX microcopy for Plumefin…"), the corrected principles, the tone dials with the humor boundary stated concretely, the three canonical exemplars as few-shot pairs (surface → copy), and output rules (character limits per surface type, sentence case, no exclamation marks in error states). Few-shot exemplars do more work than adjectives — a model imitates examples far better than it obeys "be warm."
Test-fire on one surface. Run S02 (unknown-cause transfer failure) three times. Check: does the copy stay calm, avoid blaming the bank or falsely confessing fault before cause is known, and follow your resolved apology rule? Tune the system prompt until three consecutive generations are usable, then freeze it.
Generate the full kit. One chat, frozen system prompt, all ten surfaces with their constraints stated (e.g., "S04: dialog title max 40 chars, body max 120, two buttons"). For S07 request exactly three chips under 25 characters each. Collect everything into your kit sheet, verbatim.
Rule on every string: keep / edit / rewrite. One row per string with a one-line rationale tied to a named principle or dial. Keep = ships as generated. Edit = right idea, wrong execution (fix it inline, show both versions). Rewrite = wrong idea (write yours, and note what the prompt was missing). Target honesty over pride: a kit with zero rewrites means you graded softly.
Apply extra scrutiny to the four AI genres. These carry failure modes ordinary microcopy does not. S08: does the onboarding overpromise (watch for "always", "instantly", "knows your finances")? S09: does the hedge state why uncertainty exists and what to do, or is it a vague "I might be wrong"? S10: does the refusal explain the boundary and offer a safe alternative, or does it blame the user or hide behind "I'm just an AI"? S07: do the chips suggest questions Coach can actually answer well, or do they write cheques S09 and S10 will have to bounce?
Check the kit against the money rules. Sweep every string for principle 6 violations (rounded or missing numbers where a decision happens — S04 and S06 are the danger zones) and the humor boundary (anything playful within two taps of money movement — S03, S04, S05).
Package the engine for a second writer. Final deliverable: the corrected guide (with a note on the contradiction you resolved), the frozen system prompt, the ruled copy kit, and five lines of "how to extend this" instructions — how to add a surface, when to re-test-fire, and the standing rule that every generated string gets a ruling before it ships.
The system prompt is a hiring decision, not a spell — you are onboarding a junior writer who never tires and never pushes back, which means the review discipline is the part of the system that thinks.
A voice system is only as coherent as the guide it encodes — audit the constitution first, teach with exemplars rather than adjectives, and let the keep/edit/rewrite log be the quality gate that scales you instead of replacing you.