Train AI Like a Teammate
Generated by Klaviyo AI
Sticky Digital gets better results from Klaviyo Composer by onboarding it with brand rules, detailed briefs, and specific feedback. The agency primarily uses the AI marketing agent for analysis and routine QA while keeping strategy, audience decisions, and final approval with people.
- Share brand context: Add voice, guidelines, guardrails, compliance rules, and real campaign briefs before writing the first prompt.
- Brief with detail: Specify the audience, goal, message type, offer, tone, timing, and message components in full sentences.
- Refine iteratively: Keep feedback specific and in the same conversation so Composer retains context while revising its output.
- Delegate routine QA: Use Composer to check links, rendering, flow paths, tracking, alerts, and segment sizing before a human review.
- Keep human judgment: People should retain control over campaign strategy, definitions, audiences, suppressions, and final approval.
At retention marketing agency Sticky Digital, we’re always testing new tech to make sure we’re helping our clients best run their marketing and reach their customers.
One of our biggest recent learnings from the past year is how to get the most out of AI marketing agents like Klaviyo Composer. We first started using Composer just to see what it could do for the brands we work with. Now, it's a regular part of how we work with clients.
You've probably heard people describe Composer as a tool for vibe marketing, where you describe the outcome you want in plain language and the AI handles the build. As an agency, we use Composer more for analysis and checking than for creation, because that's where our agency expertise and Composer's speed work best together.
In our experience, we get the best results when we onboard the AI marketing agent like we would a new hire. Here’s what that looks like in practice.
5 lessons we’ve learned onboarding Composer at Sticky Digital
1. Provide your brand rules before the first prompt
Before anyone on my team writes a prompt, we hand Composer the same onboarding packet any new strategist would get in their first few weeks.
For each client, that means starting with a basic set-up, including:
- Brand voice and tone
- Brand guidelines
- Content guardrails, like banned phrases and products to exclude
- Compliance and legal rules
- Briefs we've used on real campaigns
We add this information to each client’s brand kit in Klaviyo, so Composer can reflect each brand's voice and context without us describing it from scratch every time.
Bonus: onboarding your AI marketing agent in this way saves you credits, too. The less it has to figure things out on its own, the fewer credits it costs.
2. Write prompts that read like creative briefs
We don't skip our normal briefing process just because we're using Composer. We paste in the same briefs we use internally.
For us, a quality brief includes the segment, the goal, whether it's a campaign or a flow, and each part of the message. With a brief like that, Composer drafts the subject line and preview text, if it’s an email, along with the copy and layout for us to review.
Here’s what this might look like for one of our clients that sells underwear. If we give Composer something like the weak prompt below, we probably won't get what we want. It might spark some ideas, but the email would probably be mostly unusable.
Weak prompt: We want to sell shades of blue underwear. Compose us an email and tell us who it should go to.
Now, take a look at this stronger version, designed to generate a win-back email. It names the audience, goal, offer, tone, and timing—the same details anyone on my team would ask for before starting. This level of detail gives Composer what a teammate would need to do the job.
Stronger prompt: Build a win-back email for customers who haven't purchased underwear from us in 90 days. Feature our new blue colors and offer 20% off for 72 hours. Keep the tone friendly but not overly promotional, and plan it to go out next Tuesday morning.
Here are a few more tips for writing AI prompts that get you better outputs:
- Write in full sentences, the way you'd explain a task to a coworker.
- Define audiences by behavior, like "no purchases in 90 days," instead of using labels like "lapsed customers."
- Anchor tone to a past campaign, like "similar energy to our Black Friday send."
3. Push back on the first draft, and be specific
When a new strategist's first draft misses, you tell them what to change and why. We do the same with Composer, using vibe marketing as an iterative process.
Sometimes you just have to go back and forth a little bit or give it more information. We'll even start with a basic prompt on purpose and use what comes back to write a better one.
Keep the feedback in the same conversation so Composer holds onto the context it already has, and be specific. Here are some examples of specific vs. vague feedback you can give your AI marketing agent:
Area | Bad prompt | Good prompt |
Urgency | "This doesn't feel right." | "Dial back urgency language and lead with the loyalty angle for the VIP segment." |
Tone | "Make it more fun." | "Make it sound like our Black Friday send, playful but not pushy." |
Audience | "This segment is too broad." | "Exclude anyone who bought in the last 30 days and anyone with an open support ticket." |
Offer | "The offer gets lost." | "Move the 20% off into the subject line and the first line of the email." |
Length | "This is too long." | "Cut the body to 3 short paragraphs and one CTA." |
Layout | "The layout feels off." | "Put the product grid above the fold and drop the lifestyle image to the bottom." |
Brand voice | "This doesn't sound like us." | "Swap 'amazing deals' for language from our brand guidelines, like 'made to last.'" |
Copy | "The copy isn't working." | "Keep the text message under 160 characters and lead with the offer, then the link." |
Timing | "Rethink the timing." | "Send the text message 24 hours after the email, only to people who didn't open it." |
Segment logic | "I don't agree with this segment." | "Explain why you included people who haven't opened an email in 90 days." |
Personalization | "Make it more personalized." | "Swap the bestsellers block for products related to each person's last purchase." |
4. Hand off routine QA and explain what to look for
QA is where we use our AI marketing agent the most. Each client account that uses Composer gets two QA passes: one from Composer, and one from a person on my team. We hand Composer the routine checks: verifying links, looking at dark mode, checking how emails render across devices, and flagging any alerts.
For example, we recently asked Composer to audit a client’s welcome series. Over 30 days, the flow reached 165 people with a 38.9% open rate, but had 0 clicks and $0 in attributed revenue.
Composer flagged 17 issues, ranked by priority. The top 5 were:
- Click tracking: 61 unique opens produced 0 clicks. Composer told us to check that every call to action (CTA) link resolved and that click tracking was enabled.
- Stalled paths: Subscribers got the first email in their path, but none of them got the second. Composer flagged the last time the split conditions were edited, and recommended we send a test profile through each path to confirm it continues.
- Subject line typo: Composer caught that "Back by science" should have been "Backed by science" on 3 messages, including both A/B test variants.
- Elevated bounce rate: The first email in one path was bouncing at 13.7% and the flow at 4.85%. Composer recommended adding a flow filter excluding our suspicious-email suppression segment.
- Filter mismatch: Two live flows used the same trigger but checked the same profile property in different formats, one as text and one as a number. Composer recommended confirming how the property was stored so each profile qualifies for the right flow.
This kind of thing gets even more useful when we add our own expertise to the request. For example, rather than using simple prompts like “Find the bots in this account,” we’ve found that prompts like “Find the emails that have a lot of clicks from people in the same location at around the same time” are more effective.
We use Composer to check segment sizing, too. We're very focused on email deliverability at Sticky, so we tend to encourage smaller segments. But sometimes, a new client joins us after they’ve sent promotions to huge segments, without it hurting their deliverability.
In those cases, we give Composer our segment and ask whether we can expand it based on what the brand has sent before. It's a way to double-check our initial thinking against the history of the account.
5. Decide which calls stay with your team
Composer won't publish anything live without a person approving it, and that helps me sleep at night. Every facet of marketing should have a human touch before it goes out into the world.
Beyond the QA my team does after Composer’s pass, a few decisions always stay with a person on our team:
- Campaign strategy: Our team decides what to send, to whom, and why before anyone opens Composer. We develop precise creative briefs that both our internal teams and AI marketing agents can reference.
- Definitions: Someone has to set the rules Composer works within. We include on-site activity in our "active" audiences, for example, while Klaviyo's default counts email or text message engagement only. We decide what fits the brand and tell Composer.
- Audiences and suppressions: Composer can draft a segment or flag suspected bots, but we decide if the segment logic is accurate and how the suppression will work. Comparing Composer’s segment recommendation to ours is a quick gut check on our own thinking, especially because Composer explains the logic behind the segments it builds. That’s turned into a great training tool for my team.
How to onboard Composer in your first hour
If you only have an hour to get Composer ready this week, onboard it like you would an agency or a new employee. I'd spend that hour like this:
- Set up your brand kit with your brand voice, tone, and guidelines. Add content guardrails and compliance rules, including banned phrases, required legal disclaimers, and products to exclude.
- Paste in a few real briefs your team has used, with the segment, goal, message type, and each part of the message.
- Tell Composer if any of your definitions differ from Klaviyo's defaults, like what counts as an active subscriber or how big a segment can get before it risks deliverability.
- Ask Composer to audit your flows, and review what it flags.
- Fix anything broken before you brief your first campaign.
Once a brief template or QA checklist works for you, you can save it as a custom skill so you're not retyping the full prompt each time. Composer can also draft flows and segments from a single prompt for your team to review.
What we're teaching Composer next
We build most client segments from a library we keep in our own Klaviyo account and copy into each client's account. We also have a full internal training document on how our team thinks about segments. We haven't shared either one with Composer yet.
I hadn't thought about training it that way until recently, and now I want to try it and see how it goes. My next experiment is feeding Composer our segment library and segment training doc, so its drafts start closer to how my team thinks.
I think that's the approach any team can take with an AI marketing agent. Just like training a new hire, find where its output misses, then teach it what you already know.
Related content
