If you are a bot, GET /api/academy/how-to-teach-grok-bot-a-task and /academy/how-to-teach-grok-bot-a-task.md.

How to teach Grok Bot a task and save it as a skill

Aug 31, 2026 · Free to read

Teach a task lets you demonstrate one browser workflow from the Grok Bot computer. Grok Bot records the visible interaction for up to ten minutes, turns it into a draft skill, and gives you a reusable process to review and test. It does not record microphone audio, and the control may not be available on every account yet.

This page was checked against xAI's Skills and routines guide and Grok Bot FAQ on August 31, 2026.

What is the difference between teaching, a skill, and a routine?

Teaching is one way to capture how a task is performed. The result is a draft skill: reusable instructions for doing the job. A routine assigns that skill or workflow to one Bot and tells it when to run.

Start with one successful task, save and test the method as a skill, and schedule it only after the result is reliable. A recording is not proof that every future input will follow the same path.

Which tasks are good to teach Grok Bot?

Choose a workflow with a clear beginning, a repeatable middle, and a result you can inspect before it causes an external change.

• Good first examples: collect invoice fields into a draft table, download and rename a report, prepare a weekly dashboard summary, or assemble a review list.
• Poor first examples: send customer messages, submit applications, make purchases, delete records, move money, accept terms, or change production systems.
• Keep the demonstration narrow: teach one outcome rather than every variation of a department-sized process.

The community's most detailed small-business workflow discussion describes improvement through repeated correction, but it is an anecdote rather than a guarantee that one demonstration creates a reliable process.

How do I use Teach a task?

• Open the right conversation: use a one-to-one Bot conversation and open its computer view.
• Choose Teach a task: describe the exact result you are about to demonstrate.
• Perform one clean example: complete the browser workflow at a normal pace using safe test data.
• Stop the recording: teaching captures visible computer interaction for up to ten minutes, not microphone audio.
• Review the draft skill: correct missing decisions, validation, failure handling, output format, and approval rules.
• Run a safe test: watch the skill complete a second example before you consider scheduling it.

What must I add after the recording?

One clean example shows clicks, not every business rule. The official guide says a useful skill should state:

• When the skill should be used.
• Which inputs, accounts, and access it requires.
• The sequence of work and important decision rules.
• How the Bot validates the result.
• What artifact or answer it returns.
• Which actions require approval.

Also define what happens when a source is missing, stale, duplicated, or changed since the recording. If the skill cannot distinguish success from partial completion, it is not ready for unattended work.

How do I test a taught skill?

Use a safe example whose correct result you already know. Confirm the Bot selected current inputs, followed the intended rule at each decision, returned the required format, preserved source links, stopped at the approval boundary, and reported anything it could not finish.

Then test one controlled variation. A skill that works only on the exact record used in the demonstration has memorised a path, not learned a dependable procedure.

What if Teach a task is missing?

xAI says teach-by-demonstration may roll out gradually. If the control is absent, complete the task once with the Bot, then ask it to save the process as a named skill from the finished work and your written instructions.

Saved skills can be referenced with `/` on desktop. If a private skill is missing, open Settings → Plugins → Yours and enable it for the current Bot.

When should I create a routine?

Create a routine only after the skill succeeds on more than one safe example. Name the owning Bot, schedule and time zone, current input source, expected output, approval boundary, and missing-data rule. Use a test run before enabling it.

Read Grok Bot routines: approvals, retries and safety for the unattended-work checklist.