Most products wait to be opened. The user has to remember the tool exists, open it, and work out what needs doing. A proactive first message turns that round: the product looks at the user’s own data on a schedule, and writes to them only when something needs a decision, with the next step already drafted and one tap to approve it.
Two launches in the past week show the pattern clearly.
The idea and why it works
On 21 October Anthropic announced Claude Duet, with a safety update on 23 October. Anthropic says Duet sends “a short brief on the most important items across your connected apps” each morning, then “checks your connected accounts every few hours” and “messages you only when something needs your attention”. Users can “tell it what you want to hear about and how often”. Before it sends a message, makes a payment, signs a document or deletes anything, “it shows you exactly what it’s about to do and waits for your approval”. Some actions, such as trading securities or moving money between accounts, are off-limits even with approval. Duet is a waitlisted research preview, starting with US Pro and Max subscribers.
Five days later, Kiwi launched on Product Hunt (#9 of the day, 106 upvotes, free during the pilot). Kiwi describes itself as “a free AI assistant that texts you first” over iMessage or WhatsApp. It reads your calendar and email, messages you about trips, bills and plans, and says “anything that costs money or speaks for you waits for your yes”.
Neither company has published retention or engagement results, so treat this as a pattern to test on your own users. The reasoning behind it is simple. Each message is built from something that has actually happened in the user’s account, it arrives when that thing matters, and it ends in a finished task instead of a login screen. For the user, the product turns up with help at the right moment, they don’t have to go looking for it, and nothing happens until they say yes.
There are three design choices that make it work, and both launches share them:
Checks on a schedule, not constantly. A few checks a day, batched. Cheaper to run, and the product doesn’t feel like it’s watching.
Speaks only when there’s a decision to make, plus one brief. Silence is the default.
The user holds the controls. They pick topics and frequency, and every action needs their approval.
Who it’s for
Solo builders with a SaaS, app or membership whose users connect data or create records (invoices, projects, bookings, content calendars, stores).
Products where value depends on a timely action: a payment due, a client waiting on a reply, a post scheduled for tomorrow with no image.
Anyone currently running a calendar drip (“Day 3: did you know you can…”) with low click-through.
It suits products with little user data less well. If you can’t say something specific about the user’s own account, you have nothing to say first.
What you need (tools, time, cost)
Event data: your own database or product analytics, with timestamps on the records that matter.
A scheduler: a cron job, a queue or a workflow tool (n8n, Zapier, your framework’s scheduled jobs). A check every 3–6 hours is enough to start.
One channel the user already reads: email, push, Slack or SMS. Start with whichever one you already send through.
Signed one-tap action links: a link that performs one specific action for one specific record, expires, and asks for confirmation on a page before anything happens.
A preferences page: topics on/off, daily cap, quiet hours, brief on/off.
Time: about a week for the first two triggers. Cost: mostly your existing sending costs. If you use an LLM to draft the next step, budget per message and cache it.
Step-by-step
List the moments that matter. Go through support tickets and churn notes and pick the three to five situations where a user would have wanted a heads-up. Examples: invoice overdue by 7 days, client replied and is waiting, trial ends in 48 hours with setup unfinished, scheduled post missing an asset.
Write each as a trigger rule. Condition, the data you need to show, the action you’ll offer, and an urgency level. Keep a written rule per trigger so you can check it later.
Pick the default cadence. Run checks every few hours. Send at most one “needs attention” message per topic per day, at most three per week to start, and one optional daily or weekly brief that groups the rest.
Draft the action for the user. Pre-fill the reminder email, the reply or the reschedule. The user should be able to approve in one tap or edit before sending.
Ask permission and set preferences at signup. Show the topics as checkboxes, default to the two most useful, and say in plain words what data you’ll check and how often. Apple’s own guidance is a good bar: “build trust by accurately representing the urgency” of each notification, and don’t send promotional content without explicit opt-in.
Put the controls in every message. “Mute this topic”, “Fewer of these” and “Change frequency” links in the footer, each working in one click.
Retire the calendar drip for users who opt in. Keep a holdout group on the old sequence so you can compare. Our DIY holdout test walks through the set-up.
Review weekly. Cut triggers with low action rates or high mute rates, and add the next one from step 1.
Template / copy / example
A nudge for an invoicing tool:
Subject: Acme Studio’s invoice is 7 days overdue
Hi Sam, invoice #1042 to Acme Studio ($1,800) was due on 20 October and hasn’t been paid.
I’ve drafted a polite reminder that includes the payment link: “Hi Jo, a quick nudge on invoice #1042, due 20 Oct. Here’s the payment link if it’s easier: [link]. Thanks!”
[Send this reminder] · [Edit first] · [Not now]
Nothing is sent until you tap Send. You get these when an invoice is 7+ days overdue. [Mute overdue alerts] · [Change frequency]
The daily brief (opt-in):
Today, 3 things: Acme Studio’s invoice is 7 days overdue (reminder drafted). Two clients opened proposals yesterday. Your Q4 tax estimate is due in 9 days. [Review all]
Each message names the specific record, offers one action, says what will happen, and gives the user a way to turn it down. If a message can’t do all four, don’t send it.
Metrics to watch
Nudge→action rate: approved actions ÷ nudges sent, per trigger. This is the main number.
Mute rate: topic mutes plus unsubscribes ÷ nudges sent, per trigger.
Edit rate: how often users edit the draft before approving. High edit rates mean your draft needs work.
Time to action: minutes from send to approval.
Return visits from messages and retention against the holdout at 30 and 60 days.
ILLUSTRATIVE MODEL (made-up numbers, not results from Duet, Kiwi or anyone else): say 1,000 active users each get an average of 2 nudges a week, so 2,000 nudges. If 25% end in an approved action, that’s 500 completed tasks a week that started outside the app. If 2% lead to a mute, that’s 40 mutes. A trigger is worth keeping while its action rate is many times its mute rate; when a trigger’s mutes approach its actions, cut it or raise its threshold. Your own numbers will differ, so set the thresholds after two weeks of real data.
Pitfalls
Notification fatigue. The fastest way to lose the channel. Cap per day and per week, collapse low-urgency items into the brief, and respect quiet hours. Anthropic’s own list of Duet’s limitations says it can misjudge how urgent something is, so test your urgency rules against real cases.
Privacy and data access. Only read data the user connected to your product for this purpose, say what you check and how often, and let them switch it off. Kiwi, for example, puts statements such as “your data is never sold” on its homepage. Make your own promises specific and keep them.
Trust in actions. One wrong auto-send undoes months of goodwill. Keep a human tap on anything that sends, pays, signs or deletes, show exactly what will happen, and expire action links.
Drip emails dressed up as nudges. “You haven’t logged in for 5 days” isn’t about the user’s work. Every message needs a specific record and a useful action.
Hiding the off switch. If muting takes more than one click, people mark you as spam instead, and that hurts every message you send afterwards.
Variations
Brief only. Start with a weekly digest of three items and add real-time nudges once you know which ones get acted on.
Channel-first. Send nudges where the user already works (Slack, WhatsApp, SMS), as Kiwi does over iMessage and WhatsApp.
Upgrade moment. When a nudge needs a paid feature, offer it inside the action, which pairs well with the 30-day upsell window.
Save flow. A user who mutes everything is a churn signal. Route them to a short reason-matched check-in, like the reason-matched cancel flow.
Agent-drafted, rule-triggered. Let deterministic rules decide when to message, and use an LLM only to draft the action, so cost and timing stay predictable.
