Atelier Copy

Atelier — UX Writing & Microcopy

The fastest way to make a polished interface feel cheap is the words. A pixel-perfect form with the label "Submit", an error that says "Error: null", and an empty state that says "No data" reads as unfinished — because it is. This skill treats copy as design: every visible string is a decision, on the project's voice, doing a job.

Project memory: if ATELIER.md exists at the project root, read its Personality (the voice traits) and Register first — they set the voice this copy must hold. Missing on substantial work? Offer /atelier init (owned by the atelier router).

Suite map. atelier-ux decides the flows and which screen-states exist; this writes the words in them. atelier-components builds the markup the copy lives in (labels, errors, empty states) and calls here for the strings. atelier-harden owns the behavior of error/empty/form states and locale formatting; this owns their words. The atelier-perf-a11y anti-slop gate runs the "copy self-audit" Tell (and points here); atelier-review's CODE/COPY reviewer checks against this skill. Deep reference: references/fundamentals-deepdive.md.

The flow

1. Lock the voice (from ATELIER.md / Direction) → 2. Inventory every visible string →
2. Fix the high-stakes copy (errors, empty, destructive, primary CTAs) → 4. Labels & microcopy →
3. Terminology consistency pass → 6. Self-audit (cut AI-cute, fake-precise).

1. Lock the voice

Voice is constant; tone flexes by moment. Pull the three personality words + register from ATELIER.md (or the Direction Doc). Full method in references/voice-and-tone.md. The rule: voice is who you are (calm, expert, warm — set once); tone is how you read in context (an error is plain and reassuring; a success is brief; an upsell is light). One register per surface — don't slip between playful and formal in the same flow.

2. Inventory every visible string

List them by type — they have different jobs and rules: actions/CTAs, form labels + helper text, errors + validation, empty / loading / partial states, confirmations / destructive dialogs, tooltips / hints, notifications / toasts, onboarding lines, nav labels, alt text, footer / legal. You can't fix what you haven't enumerated.

3. Fix the high-stakes copy first

The strings that decide trust and task success (patterns in references/microcopy-patterns.md):

• Errors: what happened + why + how to fix. Human language, never a raw code or "null" alone, never blame the user. "We couldn't save — you're offline. We'll retry when you reconnect." Preserve their input (behavior is atelier-harden).
• Empty states: onboard, don't apologize. Name the value + one clear first action. Never a bare "No data" or blank panel.
• Destructive confirmations: name the consequence + the object + reversibility. "Delete 3 projects? This can't be undone." Not "Are you sure?".
• Primary CTAs: verb + object, specific. "Create account", "Save changes", "Start free trial" — not "Submit", "OK", "Click here".

4. Labels & microcopy

• Labels above inputs, always — a placeholder is not a label (it vanishes on focus, fails a11y). Use placeholders only for format examples; put requirements in persistent helper text.
• Verb-first, specific, parallel — actions are verbs; sibling actions share grammar ("Edit / Duplicate / Delete", not "Edit / Duplication / Remove item").
• Sentence case for UI text by default (easier to read than Title Case); reserve Title/UPPER for brand moments decided in direction.
• Helper > tooltip for anything required to complete a task (tooltips hide on touch and from keyboards).

5. Terminology consistency (one word per concept)

Pick one term per concept and use it everywhere — "Projects" is not sometimes "Workspaces" or "Boards". Match the user's vocabulary, not the database's ("Billing", not "SubscriptionEntity"). Keep a tiny glossary for the project; it prevents the slow drift that makes a product feel incoherent.

6. Self-audit (the anti-slop pass on words)

Re-read every visible string out loud and cut:

• AI-cute copy — forced wordplay, mock-poetic micro-meta ("Let's craft magic ✨"), fake-craftsman labels. Plain functional copy beats clever-wrong copy every time.
• Fake-precise numbers — 92%, 4.1×, 13.4 lb invented for effect. Real (from the brand), an explicitly-labelled mock, or cut.
• Duplicate CTA intents — "Get in touch" + "Contact us" + "Let's talk" = one intent → one label, reused. (This is an atelier-perf-a11y anti-slop Tell too.)
• Marketing slop — "seamless", "robust", "elevate", "empower", "unlock", "supercharge", "delve", "in today's fast-paced world". Say the specific thing instead.
• Jargon / unclear referents / broken grammar — anything a first-time user wouldn't parse.

Operating principles

• Copy is design. A polished UI with placeholder words ships as slop. Every visible string is a decision.
• Clarity over cleverness, always. The plain version that one-shots comprehension beats the witty one that needs a second read.
• Errors are a trust moment — say what happened, why, and how to fix it; never blame, never leak a stack trace.
• One voice, one term per concept — drift in either is what makes a product feel incoherent.
• Cut AI tells in words too — no fake precision, no forced whimsy, no duplicate CTAs, no buzzword filler.