Teach your agents how the work is done.
Skills are reusable instructions an agent follows for a job: how to triage a bug, roll back a deploy or chase a payment. Add one to any agent.
Answer a support ticket
Answer from what the product actually does, checked in the systems, rather than from a plausible account of it.
Chase a failed payment
Work a failed charge from the reason code the processor gave, rather than retrying blindly, and tell the customer before their access changes.
Check the data pipeline ran
Confirm the numbers people are about to use are fresh and complete, by checking freshness and volume rather than job status.
Confirm a reported outage
Check the whole path between a customer and the thing they say is down, so the reply is a fact rather than a reassurance.
Cut a release and write the notes
Tag a release from something already proven green, with notes written for the people who use it rather than the people who wrote it.
Escalate a ticket to engineering
Hand a support ticket over with everything needed to start, so it does not come back asking for the account id.
Find out why CI is failing
Read a red build down to the actual assertion or gate that failed, and say whether it is this change, the branch it sits on, or the pipeline.
Follow up an expiring trial
Reach a trial before it lapses with a message shaped by what they actually did in the product, and hand over the ones worth a person.
Handle a refund request
Establish the facts of a refund request and prepare the decision, without issuing money back on your own authority.
Investigate an error spike
Work out whether a rise in errors is a new fault, a new user, or a new monitor, and answer in that order before proposing a fix.
Keep dependencies patched
Work the vulnerability list down by real exposure rather than by severity label, and land the upgrades in batches that can be reviewed.
Onboard a new customer
Take a new paying account from signup to a working setup with an owner, without the four systems disagreeing about who they are.
Point a domain at a service
Add or change DNS for a service and verify it resolves and serves before calling it done, without touching records you did not come for.
Prepare the weekly revenue update
Produce the same handful of numbers every week from the same sources, with the movements explained rather than just reported.
Publish a post and announce it
Take a finished draft to live, checked in the order that catches the mistakes people see, then announce it once in each place.
Qualify an inbound lead
Find out in five minutes whether an inbound signup is worth a person's time, and write down the reason either way.
Reconcile the month's spend
Match what was spent against what was recorded, and produce a short list of the differences with a reason for each.
Reproduce a bug from a support ticket
Get from a customer's description to a failing case you can watch, using their actual request rather than an approximation of it.
Research a competitor's public pages
Pull what a competitor publishes into a comparison of facts with dates attached, using only public sources.
Restore a backup into a scratch database
Get a copy of production data somewhere safe to query, without ever pointing a restore at a live database.
Review a pull request
Read a change for the failures that matter in this codebase and leave a review somebody can act on without a second round.
Review who has access to what
Produce a list of every person with access to each system and what it gives them, with the accounts that should have been removed named.
Roll back a bad deploy
Put the last known-good build back in front of users and prove it is serving, before spending any time on why the new one broke.
Rotate a leaked credential
Replace a secret that has been exposed and prove the old one is dead, in an order that never leaves the service without a working key.
Run the weekly SEO check
Look at search performance for the four things that actually move, and report changes with causes rather than a dashboard screenshot.
Run the weekly team digest
Assemble what actually shipped, what moved, and what is stuck, from the systems that recorded it rather than from what people remember.
Set up a new hire's accounts
Give somebody everything they need on their first morning and nothing they do not, from a role rather than from a copy of a colleague.
Triage a new bug report
Turn a report into a ticket that can be worked without asking the reporter anything else, or close it with a reason.
Triage a production alert
Turn a pager alert into either a fix already moving or a written all-clear, in the order that finds the cause fastest.
Write the incident review
Assemble a timeline from the systems that recorded it, and a review that names causes rather than people, within a day of the all-clear.
Have a way of working worth keeping?
Write it down once as a skill and every agent you build can follow it.