The agent drafts. You sign the send.
Last week I wrote about what Composer actually is and why I think it favours small teams. The question I’ve been getting since is more practical: fine, it’s in my account now. What do I type into it, and how do I know when what it hands back is safe to send?
This is the working method. It has two halves: the brief you give the agent, and the checks you run on what comes back. Both matter more than the prompt-engineering tricks people are trading, because Composer’s weakness isn’t comprehension. It’s context. It knows what works across 193,000 brands. It knows almost nothing about what’s true in your account.
Mise en place, then the fast hands
Every working kitchen preps before service. Onions diced, stocks reduced, stations stocked — mise en place, everything in its place, done in the quiet hours so that when orders start flying the fast hands can just cook.
Composer is the fastest pair of hands your account has ever had. That’s the whole pitch, and it’s real. But fast hands in an unprepped kitchen don’t produce dinner faster. They produce mess faster.
So the first Composer skill isn’t prompting at all. It’s mise en place: an engaged segment that actually means engaged, suppression running, authentication aligned, metrics that fire when they should. Every one of those is invisible in the demo videos and decisive in production, because the agent inherits your account’s state completely and uncritically. Prep the kitchen before you let it cook.
What goes in the brief
Brief Composer the way you’d brief a competent freelancer on their first day: they’re skilled, they’re fast, and they know nothing about you yet. “Build a win-back campaign” is not a brief. This is a brief:
Campaign: win-back, email only.
Audience: segment "Lapsed 90–180d — engaged era" ONLY.
Exclude: anyone with Placed Order in the last 30 days,
anyone in "Sunset — final notice".
Offer: no discount. Lead with what's new since they left
(spring range, faster shipping). Discounts are a
last resort we haven't reached.
Voice: plain, warm, first person from the founder.
No urgency theatre, no countdown timers.
One email, not a series. I'll decide on a follow-up
after I see reply and click behaviour.
Notice what’s doing the work. The audience is a named segment, not a description the agent gets to interpret. The exclusions are explicit, because the most expensive campaign mistakes I see in audits are inclusion mistakes — the winback that hit last week’s buyers, the promo that went to the suppression list’s cousins. The offer logic states what’s not allowed, which matters because an agent trained on aggregate data will reach for the statistically average move, and the statistically average move is a discount.
A vague brief doesn’t get you a bad campaign. It gets you an average one — average for everyone, wrong for you.
Five checks before a draft sends
Composer drafts; you approve. That approval step is where the value lives, so treat it as a job, not a formality. These are the five things I’d check on every draft, in order:
1. Open the segment it built. Don’t trust the name — read the definition. What’s the engagement window? Which metrics count as engagement, and are any of them machine-inflated opens? Who’s excluded? Two minutes of reading definitions catches the majority of what could go wrong.
2. Confirm the trigger is real. For flows: does the metric it’s triggering on actually fire in your account? Plenty of accounts carry metrics that exist in the dropdown and never fire in practice — Placed Order being the classic when a payment gateway skips the webhook. An agent can’t hear silence. It will happily build on a trigger that never comes.
3. Weigh the email. How image-heavy is the draft? Is there real text for the filters to read, a working unsubscribe, a sane subject line without the emoji confetti? Composer’s templates are competent, but deliverability weight is account-specific: what a pristine domain can send, a shaky one can’t.
4. Question the incentive. If a discount appeared and you didn’t ask for one, that’s the aggregate talking. Ask what the message would say without it. Usually there’s a better answer — restock, story, a question — and your margin keeps the difference.
5. Check the maths on volume. Klaviyo’s own projection is that teams will go from 3 campaigns a month to 15. Capability is not a target. Every extra send draws down the same engagement budget, and quintupled volume into an unprepped list is the old deliverability spiral at agent speed. If the calendar grows, it grows into engaged segments first.
What I still wouldn’t delegate
Some jobs stay on your station even with the fast hands hired.
The calendar’s why. Composer can fill a month with sends, but it can’t know that your founder has a story about the new supplier, that the restock lands Thursday, or that your customers are still annoyed about the shipping change. What the brand has to say this month is judgment, and briefing it is the actual marketing.
Suppression and sunset decisions. Choosing to stop mailing someone is a revenue-versus-reputation call with your domain’s history attached. An agent optimising a campaign has no incentive to shrink its own audience. You do.
Deliverability incidents. When opens crater at one provider or complaints spike, the response is diagnosis under pressure — reading postmaster signals, cutting volume, re-warming deliberately. That’s the one situation where “draft faster” makes things worse.
The take
Use Composer, and brief it like a professional. Write briefs with named segments and explicit exclusions. Run the five checks on every draft, every time — they take ten minutes and they’re the whole difference between an agent that compounds your account and one that amplifies its problems. And do the mise en place first. The kitchen doesn’t get cleaner because the new hands are fast.