Your Customer.io Agent Now Takes Custom Skills—Write These Three First

On 30 October 1935, the US Army Air Corps gathered at Wright Field outside Dayton, Ohio, to watch Boeing's new Model 299 fly. It was the most advanced aircraft anyone had built: four engines, more instruments than a pilot had ever managed at once, and enough defensive guns that a reporter had already nicknamed it the "Flying Fortress". At the controls sat Major Ployer Hill, chief of the field's flying branch. About as good as pilots got.

The 299 lifted off, pitched up sharply, stalled, and came down in flames. The investigation found nothing wrong with the aircraft. The crew had left the flight-control gust locks engaged—the clamps that stop the control surfaces moving in the wind while the plane is parked. Hill died of his injuries, and so did Boeing's test pilot.

The verdict could have been "we need better pilots". It wasn't. The man flying was already one of the best in the service. The aeroplane had simply outrun what a person could hold in their head, and the answer wasn't more training. It was the pre-flight checklist: the steps written down, run the same way every time, now required in every cockpit. (This Day in Aviation has the full account.)

That's the right way to think about the custom skills Customer.io has just added to its agent. A skill is a checklist you write once so the agent runs your process the same way every time, instead of trusting it to remember. This post covers what shipped, the catch most launch coverage skips, the three skills worth writing in your first week, and the instruction text to copy for each.

What shipped on 4 June

Customer.io turned on custom skills for its agent on 4 June 2026. A custom skill is a set of reusable instructions the agent can read during any conversation—your workflow, your brand rules, your QA checks, your naming conventions—so you stop re-explaining the same things every session. Customer.io's own documentation lists the obvious candidates: reviewing campaign drafts before launch, writing in your brand voice, auditing segment conditions for stale events or attributes, and summarising weekly broadcast performance.

You create a skill one of two ways. Prompt the agent directly ("turn what you just did into a skill"), or go to Agent > Skills, click New skill, and choose Create with Agent or Write skill instructions. Once it exists, the agent reaches for the skill on its own when it's relevant, or you can call it explicitly with a slash command like /your-skill-name.

Two things worth knowing up front. Skills don't cost anything to run: only LLM actions consume AI credits, and Customer.io's billing documentation lists the agent among the features that don't use them. If you're forecasting your credit spend, skills sit outside the meter. And a skill is not a routine. A routine runs on a schedule whether you ask or not. A skill sits idle until you call it or the agent decides it fits the task in front of you.

The catch nobody mentions: skills are per-user

Here's the part the launch notes gloss over. Skills are scoped to your individual user account, not your workspace. Customer.io says it plainly: custom skills "are not shared with other users in your workspace." So the meticulous pre-launch QA skill you write does nothing for the teammate sitting next to you. Their agent has never heard of it.

That inverts the obvious pitch. "Encode your team's process so everyone works the same way" is exactly what skills sound like they're for, and it's the one thing they don't do automatically. There's no workspace-level skill the whole team inherits.

What you can do is copy. Open a skill, copy its instructions, and hand them to a teammate to paste into their own agent. The docs spell this out as the sharing method. It works, but it's manual, and copies drift the moment someone edits theirs and forgets to tell you. So the fix for a team that wants consistency is dull and reliable. Keep the canonical version of each skill in a shared doc—Notion, a Google Doc, a repo, anywhere everyone can see it. Treat each person's installed copy as a deployment of that source, not the source itself. When the process changes, you update the doc and everyone re-pastes.

Prioritisation matters more because of this scoping. You don't get unlimited skills. On an Essentials plan each user can create up to three; on Premium or Enterprise, up to ten. Run out and you delete one to make room, which frees the slot back up. Three slots per person is the real constraint that decides which skills earn a place, so spend them on the checks that catch expensive mistakes.

The three skills to write first

Pre-launch campaign QA

Write this one first, because it's the skill that stops a bad send. The job is simple: before anything goes live, the agent runs the same review you'd run yourself if you had the time and never got distracted. Customer.io's docs include a worked example you can adapt almost verbatim:

When I ask you to review a campaign before launch:

1. Check that every message has a clear goal and a matching call to action.
2. Check that every segment or trigger condition uses events and attributes from our current data plan.
3. Review timing, wait steps, and exit conditions for conflicts.
4. Summarise risks as launch blockers, recommended fixes, and optional improvements.

Add the checks that have burned you before. If your templates lean on Liquid, have the skill look for the patterns that silently render blank when a property goes null—the unset attribute with no fallback, the loop over an empty array. Tell it to flag any send to a segment without a suppression check. The skill is only as sharp as the mistakes you teach it to look for, so grow it from your own near-misses.

Brand voice

The second skill enforces how you sound. Re-typing "keep it warm, no exclamation marks, never say 'unlock'" into every draft request is exactly the repetition a skill removes. Write the rules once, with examples, and the agent applies them whenever it drafts copy.

Be concrete. Vague guidance produces vague output, so give it rules it can actually check against and a banned-words list it can match:

When you write or edit any customer-facing copy:

- Use British spelling (organise, personalise, colour).
- Sentence-case subject lines. No title case, no full stops.
- Never use: "unlock", "supercharge", "seamless", "leverage", "game-changer".
- Match the tone of these approved lines: [paste two or three real subject lines].
- Read back any line over 12 words and shorten it.

The examples do the heavy lifting. A model copying three subject lines you actually like will outperform a paragraph of adjectives describing your "voice" every time.

Segment audit

The third skill keeps your targeting honest. Segments rot quietly: an event you stopped sending, an attribute you renamed, a condition written for a data model you've since changed. None of it errors. The segment just keeps resolving to the wrong people. Customer.io names this one in its own example list, auditing segment conditions for stale events or attributes, and it pairs well with a deliberate approach to advanced segmentation.

When I ask you to audit a segment:

1. List every event and attribute the segment's conditions reference.
2. Flag any event that hasn't been received in the last 90 days.
3. Flag any attribute that is null for more than half the matching profiles.
4. Tell me the current match count and what it would be if each stale condition were removed.

Run it before you reuse an old segment in a new campaign. The five minutes it takes beats discovering after send that a third of the audience qualified on a typo.

Writing skills that actually work

Keep them lean. The hard limit is 1 MB, but that's not the limit that matters. The agent spends context reading every skill it pulls in, so a bloated skill leaves less room for your actual conversation, your workspace data, and the task at hand. Customer.io's own guidance is to keep skills "much smaller" than the cap. One skill, one job. If a skill is trying to do QA and brand voice and reporting, split it.

The name and description aren't decoration. The agent uses them to decide when a skill applies, so "Pre-launch campaign QA, run before any campaign goes live" triggers far more reliably than "Quality stuff". Write direct instructions—"Check that…", "Flag any…", "Ask me before…"—rather than broad goals, and include examples whenever you need the agent to match a format or a tone.

Two rules round it out. Don't put secrets, API keys, or sensitive customer data in skill content; a skill is instructions, not a vault. And revisit your skills when your process changes. A QA skill that checks against last year's data plan is worse than no skill, because it gives you false confidence.

What not to delegate to a skill

A skill instructs the agent. It doesn't enforce anything. That distinction decides what belongs in a skill and what doesn't.

What the agent can and can't do is set out plainly. It works within your permissions, asks before high-impact or destructive actions, and by default can't touch live campaigns or send live messages on its own. Anything it builds still goes through your normal approval and sending flow. So a skill can run a pre-launch review, but it can't be the thing that authorises the launch. The human approval step stays human.

If you want hard limits rather than instructions, that's a different layer. The guardrails that stop the agent shipping to your suppressed list and the discipline around giving it write access are what enforce safety. A skill rides on top of those guardrails; it doesn't replace them. Write skills to make good work repeatable, and keep the enforcement where it belongs—in settings and approvals, not in a paragraph of instructions you're hoping the agent follows.

Frequently asked questions

How do I create a custom skill for the Customer.io agent?

Two ways, both on Customer.io's custom skills page. You can prompt the agent to turn a task it just did well into a reusable skill. Or go to Agent > Skills, click New skill, and pick Create with Agent (the agent helps you draft it) or Write skill instructions (you fill in the name, description and instructions yourself).

How many custom skills can I create in Customer.io?

It depends on your plan, and the limit is per user. On an Essentials plan each user can create up to three skills; on a Premium or Enterprise plan, up to ten. If you hit the limit, delete a skill you no longer use and that slot frees up again.

Are Customer.io agent skills shared with my whole workspace?

No. Skills are scoped to your individual user account and aren't shared with other users in your workspace. There's no workspace-level skill the team inherits automatically. To share one, you open the skill, copy its instructions, and give them to a teammate to paste into their own agent. For team consistency, keep the master version in a shared doc and treat each person's copy as a deployment of it.

Do custom skills consume AI credits?

No. Only LLM actions consume AI credits in Customer.io, and the billing documentation lists the agent among the features that don't use credits, alongside the segment builder, content analysis and in-app suggestions. Since skills are instructions the agent reads, calling one doesn't spend credits.

What's the difference between a custom skill and a routine?

Timing. A skill sits idle until you call it or the agent judges it relevant to your current task. A routine runs automatically on a recurring schedule. Use a skill for something you want on demand, like a pre-launch review; use a routine for something that should happen regularly without you prompting it, like a Monday performance summary.

How do I call a custom skill?

The agent will reach for a skill on its own when it fits the task, but you can also call one explicitly. Use the skill's name, or type a slash command in the form /your-skill-name followed by your prompt.

What should I not put in a custom skill?

Secrets. Customer.io's guidance is to avoid adding API keys, passwords, or sensitive customer data to skill content. A skill is a set of instructions the agent reads, not secure storage, so keep credentials and personal data out of it.

How large can a custom skill be?

A skill must be smaller than 1 MB, but in practice you want it far smaller. The agent reads every skill it uses into its working context, so a long skill crowds out room for your conversation and workspace data. Keep each skill focused on one job and as short as it can be while still doing that job.

Can a custom skill stop the agent from doing something?

Not on its own. A skill instructs the agent; it doesn't enforce limits. The agent already works within your permissions, asks before high-impact actions, and can't edit live data or send live messages by default. For hard limits you rely on those settings and your approval flow, not on a skill. Think of a skill as a checklist, not a lock.

How do I share a skill with a teammate?

Open the skill you want to share, copy its instructions, and send them to your teammate to paste into a new skill on their own account. That's the documented method, because skills don't sync across users. Anyone who wants the skill has to hold their own copy, which is why a shared source-of-truth doc is worth keeping once more than one person relies on the same skill.

What happens when I delete a skill?

Deleting a skill removes it permanently and frees up one slot on your plan, so you can create another in its place. There's no shared copy left behind for teammates, since each person holds their own, so deleting yours doesn't affect anyone else's.

Can the agent write a skill for me?

Yes. If the agent completes a task well and you want to reuse the process, you can ask it to create a skill from what it just did. The Create with Agent option starts a guided conversation where it helps you shape the name, description and instructions. You can edit the result afterwards from Agent > Skills.

Sources

• Custom skills. Customer.io Docs, updated 8 June 2026.
• Create custom skills for the agent. Customer.io Release Notes, 4 June 2026.
• Ask the agent. Customer.io Docs, updated 8 June 2026.
• AI credits. Customer.io Docs, updated 8 June 2026.
• 30 October 1935. This Day in Aviation.