Onboarding copy fails on length, not on talent. Median Day-1 retention across mobile apps sits near 26%, and the global onboarding-completion rate 30 days out is only about 8% (Business of Apps / Adjust, 2025-2026). Every screen gets roughly 2 seconds before a swipe or back tap, and each step past the optimal 3-5 can shave 10-15% off completion. The 12 prompts below force tight, action-led copy on every surface — welcome, permission ask, empty state, tooltip, aha pitch, activation reminder — and cover the surfaces most teams forget: error, accessibility, locale, and re-onboarding after a major update.
TL;DR
- Keep flows to 3-5 screens; first meaningful action should land inside 60 seconds.
- Prime permission asks in-app first — soft-ask priming lifts opt-in from a ~30-44% baseline to ~65% (Pushwoosh / Plotline, 2025).
- For voice-matched microcopy, Claude Opus 4.7 or Sonnet 4.6 needs the least editing; GPT-5.5 generates more concept variants per pass. Both are current as of June 2026.
- Replace each
[bracketed placeholder]with your real product detail before sending, then A/B test the variants the model returns.
Best for
- New mobile apps and SaaS onboarding flows
- Re-onboarding after major version releases
- Permission-ask copy review and rewrite
- Activation push and email sequences
- Localization audit before launching a new market
Which model to run these on
These prompts are model-agnostic, but voice and edit-cost differ:
| Model (June 2026) | Strength for onboarding copy | Plan |
|---|---|---|
| Claude Sonnet 4.6 | Cleanest brand-voice match, lowest edit cost; cheap enough to batch all 12 | Claude Free / Pro $20 |
| Claude Opus 4.7 | Best for nuanced tone (aha pitch, upgrade nudge) | Claude Pro $20 / Max $100+ |
| GPT-5.5 | Most concept variants per pass; strong CTA/format patterns | ChatGPT Plus $20 |
| Gemini 3.1 Pro | Solid for the locale-audit prompt (multilingual range) | Google AI Pro $19.99 |
Run a single template, pick the strongest variant, then ask for 3 more in that direction rather than generating 20 cold.
1. Welcome screen variants
App: [one-line description]. First-user pain it solves: [pain]. Write 5 welcome-screen variants. Each: headline (<= 8 words), 1 sub-line (<= 14 words), 1 CTA (<= 3 words). No "Welcome aboard!", no "Let's get started!". Mark which variant is best for organic install vs paid install.2. Permission-ask copy
I need to ask for [permission: notifications / location / mic / contacts]. Write 4 ask variants. Each explains the user-value in 1 honest sentence (<= 18 words) — not the developer-value. Include 1 line of microcopy to show BEFORE the system dialog (a soft-ask primer) that previews why we are asking, so the system permission popup is not a surprise.Trigger the system dialog after a value moment, not on first launch — contextual timing can nearly triple opt-ins.
3. Empty state copy
For the empty state of [screen], write 5 variants. Each must include: 1 line on what is here (or what is missing), 1 line on the smallest first action, 1 line of micro-encouragement. <= 25 words total per variant. Forbid "no items yet" alone — always pair it with the action.4. Tooltip / coach mark
For new feature [feature], write 5 tooltip copy variants (<= 15 words each). The tooltip must explain the value, not the mechanics. Banned phrases: "Click here", "Tap to", "This is where you can". Goal: the user understands the OUTCOME of tapping it before they tap.5. Aha-moment pitch screen
My app's aha moment is when the user [action] for the first time. Write the pitch screen copy: headline (<= 7 words), 2 lines of body (<= 20 words each), 1 CTA. Voice: confident, not hyped. End with a single line previewing what comes next.6. Activation reminder push
The user signed up but has not done [key action] within 48 hours. Write 3 push-notification copy variants nudging them back. Title <= 25 chars, body <= 80 chars. Voice: not guilty, not breathless. One variant must reference a specific value the user is missing, not a generic "come back".7. Multi-step onboarding flow copy
For a 5-step onboarding flow, write copy for each step: step title (<= 6 words), sub-line (<= 18 words), primary CTA, secondary "skip for now" CTA. Steps in order: [list]. Each step should make the next step feel inevitable, not optional — but skip must still be present and honest. Keep the whole flow to 5 screens max.8. Error state during onboarding
The user hits [error] during onboarding. Write copy that does three things in <= 40 words: (a) name what happened in plain language, (b) tell them the single best next action, (c) offer 1 alternative path. No blame, no jargon, no "Oops!".9. Accessibility / VoiceOver labels
Below are my 5 onboarding screen headlines and CTAs. Write VoiceOver labels for each interactive element. Labels must be <= 30 chars, action-led ("Continue to step 2", not "Continue"), and read naturally when spoken aloud. Flag any visual-only cue that lacks a text equivalent.
[paste headlines + CTAs]10. Locale audit — adapt copy for a new market
My US-English onboarding copy is below. Adapt it for [market: UK / DE / JP / BR]. Adjust: idioms, formality level, length budget (DE tends longer, JP shorter sentences), unit references, holiday and cultural references. Flag any line that loses meaning when translated.
[paste US copy]11. Re-onboarding after major version
We shipped a major update with [N] new features. Write a 3-screen re-onboarding flow for existing users (not new installs). Tone: respectful — they already know the app. Each screen: headline, sub-line, CTA, plus an explicit "What's gone" callout if any old feature was removed.12. Free-to-paid upgrade nudge inside onboarding
At the end of onboarding I want to soft-pitch the paid tier without killing activation. Write 3 nudge variants: (a) value-led — names the single best paid feature, (b) deferred — "stay on free, we'll show you later when X", (c) trial-led — 7-day trial CTA. Each <= 40 words. None should feel like a paywall.Common mistakes
- Long welcome screens that explain the entire app instead of pitching one outcome.
- Permission asks with no value pitch, so the system popup feels like an ambush — and the 30%-to-65% opt-in gap is exactly what priming closes.
- Tooltips that say “Click here” or describe mechanics rather than the outcome.
- Skipping the error and accessibility surfaces — they fail the launch review and the App Store accessibility check.
- Re-onboarding existing users as if they are new installs, wasting the patience of people who already know the app.
FAQ
How many onboarding screens should I write copy for? Three to five. Completion drops measurably with each screen past five, and the first meaningful action should happen inside 60 seconds. Use prompt 7 but cap the flow at five steps.
Where should I ask for notification permission? Not on first launch. Use prompt 2’s soft-ask primer before the system dialog, and fire the system prompt right after a value moment (first save, first result). Priming plus contextual timing moves opt-in from a ~30-44% baseline toward ~65%.
Should I run these on ChatGPT or Claude? Claude Sonnet 4.6 or Opus 4.7 produce the most voice-consistent microcopy with the least editing; GPT-5.5 is better when you want many raw concept variants to choose from. All are current as of June 2026 — see the model table above.
How do I keep the AI from writing generic startup copy? Replace every bracketed placeholder with a concrete detail, then add your two or three banned phrases (the templates ship with the usual ones). The model copies the tightness you give it.
How do I A/B test the variants? Ship two welcome or aha variants behind a feature flag, split traffic 50/50, and read activation rate (did they reach the aha action), not raw clicks. Tools like Appcues, Pendo, or Plotline can route the split without a release.