Grok Bot: from assignment to finished work

Grok that takes a job and works on it#

Grok Bot is the part of the Grok ecosystem designed to carry out work across applications. A useful assignment describes a finished result: a comparison spreadsheet with source links, for example, rather than a general question about prices.

A Bot can have an ongoing role, such as researcher or coordinator. This chapter helps you start with one small task, review the result and gradually move to recurring work.

1. Define the resultWhat should be finished and which inputs to use.
2. Grant accessOnly to the services the task needs.
3. The Bot worksFollow its progress and questions.
4. Review the outputCheck facts and completed changes.

Scope: product steps follow official documentation checked on 23 October 2026. The assignments and practice material below are original examples. This is not a record of testing a signed-in Slovak account or a promise that every task will succeed.

When to use chat, a Bot or the API#

Choose the environment that fits the work. These are practical starting points, not an exhaustive feature comparison.

A good first project: prepare something you can check quickly. If you cannot assess the result, narrow the task or first learn what a correct result should contain.

First launch and connecting tools#

Use the official downloads at x.ai/bot. Documentation lists macOS, Windows, Linux, iOS and Android. Check your plan’s eligibility first; installing the app does not establish access.

1. Install and sign in. The guide describes Cursor sign-in and account linking for legacy SuperGrok. Follow the current screen for your subscription.
2. Start with your first Bot. On first use, click through a short intro; Grok Bot then sets up your first Bot, named Grok Bot, and opens a chat with it. Give it one main job in your first message. To rename it, click its name at the top of the chat. Add more Bots later with New → Create new Bot.
3. Start without integrations. Paste the practice text in the next section. This lets you assess a first result without granting email or calendar access.
4. Connect a service for a specific task. Find supported plugins through Connect apps. Review permissions and select the right account.
5. Take over website sign-in yourself. When asked, choose Take over, complete authentication on the Bot’s computer, then choose I’m done. Keep passwords and one-time codes out of the prompt.

Control labels can change between versions. Consult the official getting-started guide and computer and apps guide if your screen differs.

First task: turn notes into a file you can hand off#

This example uses fictional notes. Copy the full assignment into your Bot. It only needs to prepare a file; access to your accounts is unnecessary.

Turn the following practice notes into tasks.csv with these columns: Task,Owner,Due date,Status,Clarification needed. Attach it to this conversation.

Notes:
- The editor will prepare the cover design by 27 October 2026.
- Proofreading is due by 28 October 2026; no owner has been assigned.
- The budget-approval task is still incomplete. No owner or decision deadline has been set.

Do not invent missing names or dates. Use “unspecified” for missing values. Add a concrete clarification question for every gap. Do not create calendar events or send anything to other people.

Finish by stating how many tasks you found, what is missing and where the output file is.

Check the result: the file should contain three tasks. The second has no owner; the third has no deadline. The Bot must not change “not approved” to “approved”. Open the attachment and check its actual contents, not just the accompanying message.

If something is wrong, name the specific correction:

You assigned an owner to the second task, but the notes do not name one. Replace it with “unspecified”, preserve the other rows and attach the corrected file. Briefly explain the change.

This is a practice exercise with expected checkpoints, not a measured product result. CSV is deliberately simple: it makes mistakes visible before you hand over more extensive work.

An assignment that leaves less to guess#

For work inside applications, specify permissions and review criteria as well as the goal. Adapt this original template; replace the brackets with concrete details.

Outcome: [what should be finished]
Sources: [specific files, URLs or source application]
Scope: [date range, folder, number of items]
You may: [read the selected data and create a new working output]
Before my approval, do not: [send, publish, buy, delete or change original data]
Deliverable: [file type, columns, length, evidence links]
Acceptance criteria: [how I will assess correctness]
If information is missing: mark the gap; do not invent a replacement.
Budget: do not use a paid service or make purchases without approval. Process a small sample first.
Finish by separating completed actions, proposals awaiting approval and unresolved items.

Two separate controls: a prompt describes your intent; application settings also determine actual permissions and billing. Writing “stay within budget” does not create a technical spending cap. See usage and approvals.

Four useful assignments to copy#

These are original example assignments. Supply your own material and start with a small scope. A particular service requires an available connection and appropriate permissions.

1. Research you can make a decision from

Useful for comparing tools, services or public offers. Supply three specific URLs first.

Compare the offers on these three official pages: [URL 1], [URL 2], [URL 3]. My need is [specific requirement].
Return a table: provider, suitability, price and currency, billing period, main limitation, direct source, time checked.
Mark missing prices “not published”. Distinguish monthly from annual billing. Do not create accounts or buy anything. Recommend at most two options and explain what still needs checking.

2. Email drafts ready for your review

Connect the correct mailbox first. Choose a narrow label or folder.

In the connected mailbox, review at most 10 messages labelled [label] from the last 7 days. Group them into reply needed, my decision needed and information only.
For at most three messages, prepare drafts in this conversation: recipient, subject, body and a link to the original. Do not send, delete, archive or change labels. If a reply needs information you do not have, write a question for me.

3. Compare two spreadsheets

Attach two practice copies. Name the column that identifies matching items.

I attached old.csv and new.csv. Compare them using the ID column. Create changes.csv with columns ID,change_type,column,old_value,new_value.
List added, removed and changed items separately. Do not silently merge duplicate or blank IDs: report them as issues. Preserve the original files. Finish with the row count of each input and the number of differences found.

4. Review a website before publishing

Provide a public or accessible preview. Specify the pages and screen widths.

Review this website preview: [URL]. Visit the homepage and these three pages: [URLs]. Check navigation, language switching, links and readability at widths of 390 and 1365 pixels.
Return at most 8 concrete findings: page, reproduction steps, expected and actual behaviour, severity and a screenshot. Separate verified defects from improvement suggestions. Do not deploy anything or change hosting settings. Mark anything you could not verify.

Review questions: Did the Bot open the cited source? Does the attachment contain every row? Does each link support the specific claim? Is “done” backed by an output you can inspect yourself?

From a good example to recurring work#

For your first automation, we recommend public sources and results delivered only to the conversation. Request and refine a one-off result first. Then adapt this assignment:

Save our verified news-checking process as a skill called “Weekly changes”. Propose a routine for this Bot every Monday at 08:30 in Europe/Bratislava.
Sources: [three specific official URLs]. Track only changes since the last successful check.
Output in this conversation: at most five material updates, each with its date, direct link and explanation of its impact. If nothing material changed, do not invent an update.
If a source is unavailable, mark it and do not present old information as new. Do not send anything elsewhere or modify the websites.
Show me the schedule, time zone and proposed next run. Do not create the routine yet. Create it after my approval, then confirm its state.

Before enabling it, check: the URLs, Bratislava time zone, output destination and missing-data behaviour. After its first real run, check for repeated items. Pause the routine when you want a break; closing the app does not tell it to stop.

If Teach a task is available, you can demonstrate a workflow. Review the resulting procedure anyway. The mobile guide says routine schedule/instruction edits and test runs require desktop.

What belongs in a saved skill

State when it applies, required inputs, steps, validation and the output. Include what to do when a source is missing. The skills documentation describes private skills as shared across your Bots; another Bot still needs the relevant connection or login.

Review the saved skill before its next use. List its inputs, steps, result checks, output destination and actions requiring my approval. Compare it with the last successful run. Do not run or overwrite it yet.

Multiple Bots, projects and handoffs#

Start with one Bot. Add another when it has a distinct ongoing role. For example:

On desktop, create a group through New by selecting two to six Bots. For approvals and sign-in, also check the individual Bot’s conversation; the group does not show every such step. Collaboration details.

Prepare one briefing on [topic] together. The researcher supplies direct sources; the writer prepares the draft; the reviewer checks each factual claim against its source. The writer applies the review and delivers one final version and unresolved questions. At every handoff, identify the current files. Nobody publishes anything.

Three agreeing answers do not establish correctness. The reviewer should open the original source. Separate roles, names and conversations also do not create separate access boundaries on the shared computer.

Three different collaboration models: a group connects your Bots; a template gives another user a copy of the configuration; a Team Bot provides a shared, managed configuration, with members holding their own conversations. Check what access and material will become available before sharing.

Grok Bot, @bot and @grok on X: the distinction for Slovakia#

Grok Bot is the app and Bot system. @bot on X forwards a request to your Bot; you follow the work in its conversation. @grok is a different account and interaction on X. The names are not interchangeable. X’s explanation.

For a Slovak account, do not make tagging @bot under a post your first step. If you have access to the Grok Bot app, paste the link and assignment there directly. This is a separate route; it does not bypass the regional rule for tagging on X.

The X plugin provides additional tools for X content. Connecting it does not itself establish eligibility for @bot tagging. Review the requested permissions and terms in your particular account before connecting it.

Access, subscription and usage#

Before buying access or assigning a longer job:

1. Check which account is signed in and whether you have the new subscription or a legacy linked plan.
2. Review remaining usage, renewal time and additional-usage settings. Look for Usage & Billing in the app; The X Subscription also shows usage on the Grok dashboard.
3. Measure a small representative task first. A subscription does not mean unlimited tasks or include purchases from other services.

Use Plans and subscriptions for prices, taxes and account transitions. The original Bot launch article may not describe today’s billing. Linking an existing subscription and unifying accounts under The X Subscription are different steps.

What you allow and what happens to your data#

Before approving, inspect the destination, account, recipient, content and scope of the change. “Send these three reviewed drafts to these three recipients” is clearer permission than “handle my email”. Review again if the recipient or content changes.

Use sample or sanitised data for practice tasks. Connect a service only when a task needs it. If you also enable work on your own computer, assess that access separately from the Bot’s cloud computer.

When you stop using it: save results first, pause unwanted routines and review active connections. When revoking access, also check authorisation at the connected service. Deleting a conversation, disconnecting an app and deleting an account are different operations.

For company data, have the administrator approve settings under your organisation’s rules. This guide does not certify that a particular deployment meets your contractual or legal requirements.

When work stalls or the result is wrong#

To stop work immediately: send the Bot a direct Stop now instruction. This does not undo completed changes. Inspect what happened before deciding how to repair it.

Do not start a computer repair with a reset. The official troubleshooting guide starts with simpler recovery steps; resetting can lose recent unsaved work. Give support the app version, time and time zone, and exact error without passwords or private material.

When the job is actually finished#

The Bot’s closing message is only one place to check. Inspect the deliverable and the state of the service where the work happened.

1. Open the result: the file downloads and includes the promised sheets, attachments or sections.
2. Check a sample: open two sources, recalculate one total and inspect an exception.
3. Verify changes: when a service was updated, inspect the correct account, record and value. A draft is not a sent message.
4. Read the limits: missing information should be marked, not silently guessed.
5. Record the next step: who receives the work and what still needs a person.

Run a final check. Return links to finished files, the changes actually made, sources used and unresolved items. Label each item “complete”, “draft ready” or “unsuccessful”. Do not send or publish anything new.

Your first week with a Bot#

This is our suggested practice plan, not a product feature or a performance promise.

1. Day 1: create one Bot and complete the notes exercise. Check every row.
2. Day 2: try one real task using public material. Define what “done” means.
3. Day 3: add one necessary service. Review the account, permissions and approval boundary.
4. Day 4: repeat the workflow with new inputs. Identify instructions that need improving.
5. Day 5: create a routine only after the result is reliable. Check its first run, usage and pause control.

The task is done when: the output exists, opens correctly, matches its sources, covers the agreed scope and clearly identifies unfinished items. A message saying “handled” is not enough.

Use the Prompt workshop to build further assignments. Find review principles in Complex tasks.