Teach a Grok Bot the job: follow along, then routine

Grok Bot can learn a workflow from a live demonstration. You ask a Bot to follow along once through a multi-step path; it persists that path as a routine and can re-run it on demand or on a schedule. This is one of the product’s actual differentiators. It is also easy to automate a sloppy demonstration.

Teaching is not a substitute for a written brief. It is how you capture clicks, systems, and exceptions after the brief already says what “done” means.

When teaching is the right tool

Use follow-along when:

• The path crosses apps that have no clean plugin or MCP server
• The clicks matter as much as the words
• You can complete the job once yourself in a non-destructive environment
• You are willing to watch the first replay before scheduling it

Do not start here when you still cannot name the output, the stop condition, or the approval boundary. Write those first with the workflow builder, or copy a playbook.

A safe teaching loop

1. Write the contract. Outcome, sources, output format, do-not list, missing-data behavior, review gate.
2. Pick the owner Bot. One Bot owns the routine. A group chat is a later coordination tool, not the teacher.
3. Ask it to follow along. Say that you will perform the path once, that it should record the systems and decisions, and that it must not act externally during the demo.
4. Do the job once, on staging, a sample folder, or a throwaway record.
5. Ask it to restate the path as numbered steps plus exceptions, then correct the restatement.
6. Replay once on a second input. Watch Agent Computer. Inspect the artifact.
7. Save a skill for the method, then create the routine only if the replay matched the contract.

Official documentation still recommends testing a task manually before scheduling it. A demonstration is that manual test, recorded.

What to say

Follow along while I complete this job once. Record the systems I open, the clicks that matter, the files I touch, and every place I stop to judge.
Do not send, publish, purchase, or delete anything during this demonstration.
When I finish, restate the path as numbered steps, including exceptions and the approval stop.
We will replay it on a second sample before any schedule is created.

If a login, CAPTCHA, or two-factor prompt appears, take over the computer, complete only that step, and tell the Bot to continue. Do not paste the secret into chat. See Agent Computer.

What a learned path still cannot guess

Teaching captures how you moved. It does not capture:

• This week’s list, unless you put the list in the routine inputs
• A policy you never showed
• Permission to send because you drafted during the demo
• A second Bot’s private conversation

Put durable rules in the Bot profile and the skill. Put changing inputs in the routine. Keep consequential actions behind approval even after the tenth clean run.

Desktop vs iPhone

xAI’s mobile guide currently says some advanced desktop controls and teach-by-demonstration workflows are not available on iPhone. You can still start work, approve steps, and review results from the phone. Create and edit the taught routine on desktop. Details: Grok Bot on iPhone.

After it is a routine

Official documentation currently describes up to 50 routines and visibility into 20 recent run records. Check enabled state, timezone, required plugins, and recent runs before duplicating anything. Pause the routine if a source system is down; do not let it “helpfully” reuse yesterday’s file.

Sources and verification

Last verified: August 27, 2026.

• Skills and routines (xAI)
• Grok Bot overview (xAI)
• Grok Bot for iOS (xAI)
• Use cases (xAI)