This page is the end-to-end workflow for writing onboarding copy with AI. For copy-ready templates, see app onboarding copy prompts.
TL;DR
Onboarding has 3-5 screens to keep a brand-new user from bouncing, and copy does the persuading before the product has proven anything. Use AI for the mechanical wins (five headline variants, hitting a word ceiling, a non-guilt-tripping skip CTA) and feed it your real value prop, real activation action, and real permission reasons so it does not invent features. Then cut every headline you stumble over reading aloud, and instrument per-screen drop-off so you know which screen to rewrite next. As of June 2026, Claude Sonnet 4.6 is the better default for natural brand voice; GPT-5.5 is faster when you want many rough variants to choose from.
The task
Your onboarding has 3-5 screens to convince a brand-new user not to bounce. Each screen gets roughly two seconds to earn a tap, and the copy carries most of the load before the user has any real product experience. The stakes are high: the global average onboarding completion rate was about 8.4% in Q2 2025, while top-performing apps land in the 40-50% range (per Business of Apps). The copy is one of the few levers you fully control.
When AI is the right tool
AI is genuinely good at three things here:
- Producing five variants of one headline so you can A/B test tone instead of guessing.
- Compressing a sentence to the word ceiling a phone screen tolerates (roughly 7 words for a headline, 14 for a subhead).
- Writing a “skip” CTA that does not sound passive-aggressive.
Model choice matters for copy. As of June 2026, head-to-head copywriting comparisons consistently favor Claude (Sonnet 4.6 for most drafts, Opus 4.7 when brand voice has to be exact) for natural, on-brand prose, while GPT-5.5 is fast and flexible for generating many rough options to sift through. A reasonable workflow: brainstorm variants in GPT-5.5, then refine the finalists in Claude. See ChatGPT vs Claude vs Gemini for the longer comparison.
When not to rely on AI alone
AI does not know your activation metric. If a headline implies a feature you have not shipped, or asks for a permission you have not justified, you have manufactured a churn moment. Pass real product names and real permission reasons; do not let the model invent them.
The permission screen is the single most expensive place to be vague. The average iOS push opt-in rate across 600+ apps was about 56% in 2025 (Pushwoosh), ranging from ~70% in fintech down to ~23% in hypercasual games. A custom pre-permission “primer” screen that explains the value before the system prompt fires can lift opt-ins by 2-3x. AI can draft that primer, but only you know whether the value claim is honest.
What to feed the AI
- App value proposition in one sentence.
- The single user action you want them to take after onboarding (the activation event).
- The “aha moment” — the first screen where the product visibly works for them.
- Tone words (confident, warm, playful, technical) — pick two, not five.
- Number of screens.
Copy-ready prompt
Replace each [bracket] with your real input. Keep brackets out of prose; they live only inside this code block.
Write onboarding copy for a [N]-screen flow.
Value: [one-sentence value prop]
Top action: [the activation event]
Aha moment: [the screen where the product visibly works]
Tone: [two tone words]
For each screen return:
- headline (max 7 words)
- subhead (max 14 words)
- primary CTA (max 3 words, action verb)
- skip CTA (neutral, not guilt-tripping)
- one line: why this screen exists
Rules: no "Welcome!" headlines. No permission ask without a value line.
Then recommend which screen to A/B test first and why.How to check the output
Run every screen through this checklist before it ships:
| Check | Pass condition |
|---|---|
| Read aloud | Each headline lands in under two seconds; stumble = cut |
| CTA verb | Action-led (“Start tracking”, not “Continue” or “Next”) |
| Word budget | No screen requires reading more than ~25 words total |
| Permission | Every permission ask is preceded by a one-line value pitch |
| No filler | No “Welcome to …” headline wasting the first screen |
| Honesty | Every claimed feature actually exists today |
Common mistakes
- “Welcome to …” headlines that burn the most valuable screen.
- Permission asks (notifications, location) with no value pitch — the difference between a ~56% opt-in and a much lower one.
- Vague subheads that just restate the headline.
- Skip CTAs hidden in 10pt grey or worded to guilt the user. Hidden skip links feel like a trap and cost you the next session.
- Six or more screens. Appcues data shows completion drops roughly 15% for each screen beyond five.
How many screens, by app type
Use the lighter end of the range for simple apps and the heavier end only when the product genuinely needs setup.
| App type | Realistic completion target | Screens |
|---|---|---|
| Games, social, content | 70-80% | 2-3 |
| Productivity, marketplaces | 50-65% | 3-4 |
| Fintech, complex setup | 50-65% | 4-5 (with priming) |
Targets are 2025-2026 category benchmarks; your baseline matters more than the absolute number.
Next steps to keep improving
Instrument drop-off per screen. The screen with the biggest fall-off is the one where your copy is failing. Re-run the prompt with the failing screen’s current text in the input and ask for five sharper variants. In most flows the first screen and the permission screen account for the largest share of churn, so start there. For the listing copy that sits in front of onboarding, see app store listing prompts.
FAQ
- How many screens is too many? Five is the ceiling for most consumer apps; completion drops about 15% per screen beyond that (Appcues). Past five you are fighting the user’s patience, not a competing app.
- Which model should I use? As of June 2026, Claude Sonnet 4.6 for natural, on-brand copy; Opus 4.7 when voice must be exact; GPT-5.5 when you want many quick variants to choose from. Brainstorm wide, then refine the finalists.
- When do I ask for push permission? Not on a cold onboarding screen. Show a primer that explains the value, then fire Apple’s prompt after a real value moment (first save, first result). Priming can lift opt-ins 2-3x over the cold system prompt.
- Should I localize first or polish English first? Polish English to a working draft, then localize with the model. Translating broken copy gives you broken copy in five languages.
- What if my product is B2B? Raise the subhead allowance to ~20 words and add a “for teams” line. Otherwise the workflow is identical.
Related
Tags: #Workflow #Onboarding