Today we are introducing Cube Agent Skills.

When we shipped Cube Agentic Analytics, the bet was that the semantic layer could be the system of record for AI—every answer grounded in the measures, dimensions, joins, and access policies your team already defined. That works well for a one-off question. But the most valuable analysis in any company isn't a one-off. It's the report someone produces every Monday, the audit a data team runs the same way every quarter, the multi-step routine that lives in one person's head and gets re-typed as a long prompt every single time.

Skills are how you stop re-deriving that work. A skill captures a multi-step analytical workflow once, gives it a name, and lets anyone on the deployment run it—and get the same vetted result the author intended.

A skill is markdown in your data model

The thing I like most about how this shipped is that there's nothing new to learn about where skills live. A skill is a markdown file in your Cube project, under agents/skills/, right next to the rules and certified queries you already manage as code. The frontmatter carries the metadata; the body is the instructions the agent follows.

---
title: "Weekly revenue report"
description: "Use when the user asks for a weekly revenue summary or WoW trends."
---


Produce a weekly revenue report:


1. Pull revenue by week for the last 8 weeks
2. Compute week-over-week deltas
3. Call out the largest movers and a one-line summary
  • title is the user-facing label—the button text in chat.
  • description is what the agent matches a free-text request against when it decides to run a skill on its own.

The skill's name comes from the filename, so weekly-revenue-report.md becomes the weekly-revenue-report skill. You author it in the IDE, commit it, and ship it through the same dev-branch-to-master flow as everything else in your project. A skill on a dev branch is testable before it ever reaches production, and it inherits the governance you already have: code review, access policies, version history.

This is the part that's specific to Cube. A skill isn't an object you create in a separate admin UI and hope someone documented. It's data-model-as-code—reviewed, versioned, and deployed with the rest of the semantic layer.

A few more skills

The weekly report is the obvious first skill, but because a skill is just instructions, almost any routine your team re-runs by hand can become one. You can run a skill from anywhere you work with the agent—Analytics Chat, a Workbook, the IDE—and what it produces is whatever its instructions describe: an analysis, a saved dashboard, even a change to the model itself.

A churn analysis pins down the definition of "churned" so every run returns the same cohort:

---
title: "Customer churn analysis"
description: "Use when the user asks about customer churn, customer retention, customers who stopped buying, or inactive customer analysis."
---


Analyze customer churn:


1. Treat a customer as churned after 90 days with no order following a prior purchase
2. Break churn out by signup cohort and plan tier
3. Compare to the prior quarter and call out the biggest movers

The output doesn't have to be a chat answer—a skill can assemble and save a whole dashboard:

---
title: "Revenue overview dashboard"
description: "Use when the user asks to build the standard revenue overview—MRR, new vs. expansion, and revenue by region—as a dashboard."
---


Build the revenue overview dashboard:


1. Add an MRR tile with a month-over-month sparkline
2. Add new vs. expansion vs. churned MRR as a stacked bar by month
3. Add revenue by region as a map and revenue by plan as a table
4. Lay the tiles out and save it as "Revenue overview"

And a skill can change the data model itself—this one captures how your team adds to it, so a new metric lands defined, tested, and reviewed the same way every time:

---
title: "Add a metric to the model"
description: "Use when the user needs a new metric or measure defined in the data model and added to the right cube and views."
---


Add a new metric to the model:


1. Create a dev branch to work on
2. Find the right cube and check the measure isn't already defined—if it is, flag the duplicate instead of adding another
3. Add the measure with a clear name, title, and description
4. Include it in the relevant views
5. Run a short test query that confirms it returns the expected number

Three ways to run one

A skill is only useful if people actually reach it, so we surfaced them in the three places a request tends to start:

  • Buttons. The chat empty state shows one button per skill, labeled with its title. Click it and the workflow runs.
  • A slash menu. Type / in the chat input to filter skills by title and description and pick one with the keyboard. When skills are available, the input hints "Type / for skills."
  • The agent picks it. When a user's plain-language request matches a skill's description, the agent runs that skill on its own—no button, no slash. Skills work this way for both the agents you configure and the Auto agent that routes across them.

The slash menu in Analytics Chat listing available agent skills, each with a title and description

What's in v1, and what's next

This first release keeps the skill itself deliberately small: a title, a description, and an instruction body—no special syntax, no action API to learn. Every example above uses that same plain format. The only difference is what the agent does with it where it runs. That's enough to turn most repeated workflows into a shareable artifact.

What we left out is just as deliberate: per-skill data scoping, external actions that reach outside Cube (posting to Slack, hitting a webhook), parameterized inputs, and per-skill model selection. These are all directions we're looking at, but I'd rather ship the core well than ship parity with a feature list. If one of those would change how your team would use skills, tell us, and we'll weigh it for what comes next.

The bigger thing we're building toward is scheduled tasks. Today someone still has to reach for a skill—click the button, type the slash, ask the question. Pair a skill with a scheduled task and it runs on a set cadence without being asked, so the report someone produces every Monday is generated and waiting on Monday morning.

Get started

Cube Agent Skills are available now. If you're already on Cube, add a markdown file under agents/skills/ on a dev branch, commit it, and the button shows up in chat once it's deployed—start with a workflow your team already runs by hand. If you're new to Cube, request a demo and we'll walk you through the full Agentic Analytics workflow, skills included.