A Skill Is How. A Routine Is When.

xAI wrote it in their own docs. We already published it for Claude. A skill is how. A routine is when. Do not schedule a demo.

This is the glossary for the last two days of Grok Bot pages. The roster lives at /grok-bot/. The job shape — schedule, QA, working files — still lives at persistent agents. The Claude how-to is 18 scheduled tasks and how to use Claude. The figurehead install is how I use Grok Bot. Official xAI: skills, routines, and automations.

Same ladder, vendor clothes

A skill is a document. An agent is a role. A job is a schedule. Vendors named the middle two rungs after their products. Do not collect a fourth SOP for each new button.

ChatGPT still says Scheduled tasks. grok.com Build still has a 7-day /loop. That loop is not Grok Bot. Do not put production jobs on a prompt that expires.

What this looks like on this Mac (18 Aug 2026)

I did not write this page from a vendor PDF. I wrote it after a week of actually using the three runtimes, then watching the meters disagree with the marketing email.

On Grok Bot I staffed 12 named desks under one coordinator — Ops, Strategy, Training + Documentation, Article writer, Website Builder, Personal Brand Expert, Fleet monitor, Usage Monitor, IT Support, Customer support, Executive Assistant, and a private desk I do not document. That roster is the public playbook at how I use Grok Bot. I kept mixing up the colored shapes in the sidebar and wanted a taxonomy. They are faces, not a role code: what the sidebar icons mean. I even have two Executive Assistant rows with different avatars. The icon is who. The skill is the method. The routine is the clock.

A named desk without a skill is still a chat window with a photo. IT Support is the one with a public can/can’t list. Proof from this weekend: theninetriangles.com login for access@blitzmetrics.com was reset from the shared computer on 16 Aug. BlitzAdmin could not rotate that password. The public site stayed up. That job is not yet a routine. Two other desks already have clocks — we photographed them rather than inventing a stack of new ones on an empty weekly meter.

On Cursor I built this glossary, the fleet topic wheels, and the rest of this week’s factory. Cursor is the workshop. The live spending tab at 18 Aug, early afternoon Mountain:

The 7:00am mail from Cursor said I had used up Other Models and should switch to Grok 4.6. The default in this editor already is Grok 4.6. The only 100% bar is Grok Bot. That is why we saved the skill today and did not enable new routines: Extra Credits after a weekly ticket are grocery prices. I staffed a Usage Monitor desk on the Grok Bot roster for exactly this class of lie.

Claude already had the worked example of a scheduled job that watches the other cafeteria: the token-usage dashboard. Same rung. Different clothes. The chats are still rented. The files are ours: chats are rented, files are ours.

A useful skill states six things

xAI listed them. They match a Task Library skill: trigger, inputs, steps, definition of done, example, approval. If a Bot bio is the only copy, you do not have a skill. You have a costume.

1. When to use it
2. Required inputs and access
3. The sequence of work
4. How to validate the result
5. What to return
6. What requires approval

Canon stays in markdown we own: github.com/dennisyu/blitzmetrics-skills, the Task Library, Dennis OS. A Grok Bot skill is a runtime copy. Refresh from GitHub. Do not re-upload a zip as the source of truth. That is one shared brain.

A routine confirms six things

Ask the Bot that should own the recurring job. Not a swarm. One owner.

1. The owning Bot
2. The schedule and time zone — or the event matcher
3. The input source
4. The expected result
5. The approval boundary
6. What happens when the source is missing (report failure; do not reuse stale data)

Event triggers (Slack, GitHub) are Cursor account integrations. They are not the Slack or GitHub plugin. Write a narrow rule. “Every new message” is how you burn the weekly meter on noise.

A test run performs real work. Safe inputs. Keep send, spend, delete, publish, and production changes behind a yes. After a site, connector, or source format changes, test again.

Manage it: open the Bot, View conversation details, Routines. Enable, pause, test, edit, inspect, delete. Delete is immediate and has no undo. Deleting a Bot deletes its routines. Cap 50 routines per Bot. The app keeps 20 run records. If you are away a long time, Grok Bot may pause unattended usage until you answer — review paused routines when you come back.

Paste this, do not invent a fourth language

Claude already had the sentence. Grok Bot and Cursor get the same sentence in their clothes.

Claude / Cowork
Create a scheduled task: every weekday at 8:00 AM local, run my
<skill-name> skill against the current inputs. Leave the report
in this conversation. Do not send, spend, delete, or publish.
If the source is missing, report the failure. Do not use old data.

Grok Bot (owning desk)
Save the process we just used as a skill called “<name>.”
Include source systems, risk definitions, output format, and
the rule that customer contact / email / publish / spend
always requires approval.

Then: every weekday at 8:00 AM America/Denver, run that skill.
Post a linked watch list in this conversation. Do not contact
anyone. If the source is unavailable, report the failure.

If the skill is missing from /, open Settings → Plugins → Yours
and enable it for this Bot. Test run before you leave it on.

Cursor
Save this as a Cursor skill first (how). Only then offer a
Cursor Automation (when). Do not open the Automations editor
for a workflow that has not passed once on safe inputs.

Trust the same way on every runtime

• Automate preparation before execution. Draft, reconcile, recommend first.
• Approval for sending, purchasing, deleting, publishing, or changing production.
• No-data and stale-data policy. Idempotent retries. Report partial completion.
• Do not treat separate Bots as a security boundary. They share one computer.
• Passwords and 2FA stay in the secure handoff. Never in ordinary chat. Official: approvals, security, and privacy.

House rule on the hub is the same sentence: docs ship by default and we iterate. Still stop for send, money, delete, permissions, or production infra.

What we will not do

We will not keep a Grok-only copy of every Task Library skill as a Bot bio. We will not call grok.com Automations or a 7-day Build /loop the production host. We will not buy SuperGrok Heavy hoping it makes named Bots run four-at-once. We will not schedule a skill that has not passed once.

Standing SOP, 18 Aug 2026. Sources: xAI skills and routines, Grok Bot overview, approvals, persistent agents, 18 scheduled tasks, how to use Claude, Grok Bot desk, how I use Grok Bot. Receipt: how we harmonized skills and routines.