A free guide from Factwire. https://www.factwireai.com
Start here
An AI agent does work for you. It does not only answer questions. Grok Bot is one. It can use a cloud computer, connected apps, websites and files. [S5]
This guide covers five headaches. Each one has the problem in plain words, why it happens, and a fix with Grok Bot steps. The examples come from public posts and help pages. A post shows what a person said. It does not prove the event happened exactly as told.
The five headaches:
An update breaks something that worked
Limits and costs run out faster than expected
The assistant forgets
Wrong answers and false done reports
The agent does something you did not ask for
Copy-ready prompts are set apart with a bar on the left. Paste them as written.
Headache 1: An update breaks something that worked
The problem
It worked yesterday. Today it does not.
One OpenClaw user posted that an update broke all of their scheduled jobs. The only options were to go back one version or wait for the next release. [S9]
Microsoft moved personal Copilot users to an updated app. Microsoft says chats move over, and that this can take time. It also says some features are retired. Microsoft says Group Chat content did not carry forward after August 18, 2026. [S8]
Why it happens
The tools you rely on change, and you do not control when. Grok Bot help pages say websites and tool behavior change over time. [S4] They also say to re-test a routine after a website, app or source format changes. [S1]
The fix
Do the job once by hand with Grok Bot. Get it right. Then save the method as a skill. Automate it last. [S1] [S2]
❝Do this task with me once, now. When it works, save the steps as a skill. Include when to use it, what you need, how to check the result, and what needs my approval.
Use a connected app when one exists. Help pages say apps are more reliable than clicking through a website. [S2]
Give every routine a missing-data rule. Tell it to report the failure and not use old data. [S1]
❝If the source is unavailable, tell me and stop. Do not use old data.
After any change, press Test on the routine. A test run does real work, so use safe inputs and keep write actions behind approval. [S2]
Check five things on each test: it used current inputs, it met the output format, it kept a source trail, it stopped at the approval point, and it made failure states explicit. [S2]
Pause a routine when its source system or workflow changes. [S4]
Save what you cannot lose somewhere else. Microsoft told Copilot users to save Group Chat messages to a document or notes app and to download images. Its date was August 18, 2026, so treat this as a habit, not a live deadline. [S8]
Headache 2: Limits and costs run out faster than expected
The problem
The weekly limit is gone by Wednesday. Or a bill shows up that nobody planned.
One person posted that an agent re-reads the same context on every pass, so spending goes to the same files again. [S13]
Another person audited every agent routine they run and found one that was not on the list. It kept waking itself up and hit the usage limit on its own. [S14]
A third posted that a runaway loop overnight almost cost $1,200. They said provider alerts only fire after a threshold is crossed. [S10]
One Grok Bot user posted that most of their tokens went to re-reading old chat. This is one person reporting on their own setup. [S19]
Why it happens
Grok Bot usage is metered, and how fast you use it depends on the kind of work and how much effort each task takes. [S5] On The X Subscription, Grok Bot draws from the same weekly pool as Cursor and the Grok apps. [S6]
The monthly extra-usage cap does not stop a task that is already running. It can go a little past the cap. [S5]
One more mix-up. The Grok app and Grok Bot are two different apps. A paywall or limit in the Grok app is not about Grok Bot. [S5]
The fix
Look at your usage. In Grok Bot, open Settings, then Usage & Billing. [S7] On The X Subscription, usage also shows at https://grok.com/dashboard. [S5]
Set a cap before big work. On The X Subscription, leave Auto Top-up off or set a monthly additional usage limit at https://grok.com/dashboard/billing. On a legacy Cursor plan, open Settings, then Usage & Billing, then On-demand monthly limit. [S5]
Remember the cap does not stop a task mid-run. [S5]
Scope each job tightly. Check your usage before a large or long job. The help pages give this advice for the free trial. It works as a habit on any plan. [S5]
❝Before you start, tell me the steps you plan to take. Stop and ask me if the job needs more than those steps.
Review your routines on a schedule. Help pages say to review installed connectors and active routines regularly. [S4]
❝List every routine you run. For each one, give the schedule, the last run, and what it does. Do not change anything.
Keep event routines narrow. Help pages warn that a listener on every new message creates noise, uses up your usage, and acts on things that do not matter. [S2]
Try a short handoff note between jobs. Ask Grok Bot to save a few lines on decisions made, open items and next steps, then start the next job fresh from that note. One developer reports passing a handoff note under 60 lines instead of an old transcript. [S21] This is a suggestion. The savings are not measured here.
Headache 3: The assistant forgets
The problem
You told it a rule last week. It does not know the rule today.
People post about agents that forget names, forget things, and repeat tool calls that did not need repeating. [S22] [S23]
After Microsoft moved Copilot users to the updated app, Microsoft says chats and content might still be on the way. It says to restart, check the sign-in and try again later. [S8]
Why it happens
A Bot keeps stable preferences, important facts and summaries of its work. It does not replay every old message. [S3]
Help pages say memory is not a source of truth. [S3] A summary can also leave things out, so keep the original messages when it matters. [S20]
Each Bot has its own memory. Only a few facts about you, like your name and time zone, are shared across your Bots. [S5]
The fix
Say the rule out loud and ask it to keep it. Help pages give this example of a saved rule: always ask before you send an email. [S3] They say to tell each Bot its standing boundaries and ask it to remember them. [S5]
❝Remember this rule: always ask me before you send an email. Save it, then tell me exactly what you saved.
Check what it remembers. Ask the Bot what it remembers about you. [S5]
Fix stale facts directly. Tell the Bot what is wrong and what is right. [S2]
Keep changing facts in the source, such as your calendar or your files. Ask the Bot to reopen the current data before any decision that matters. [S2]
❝Before you answer, open my calendar and read it again. Do not rely on what you remember.
Keep logins, cards, addresses and phone numbers in Vault. Do not put them in memory. [S3]
For a long job, ask for a short handoff note saved as a file. [S21] Keep the original messages too. [S20]
❝Save a handoff note for this job. Put in the decisions we made, what is still open, and the next step. Keep it under 60 lines.
Save a copy of anything you cannot lose. Microsoft told Copilot users to do this with content from a retired feature. [S8]
Headache 4: Wrong answers and false done reports
The problem
The assistant says it is done. It is not. Or it gives a confident answer that is wrong.
One user of another assistant posted that when they ask for status, it says it did something but did not do it. [S15]
Another posted that their agent confirmed every post from its own session. They opened the same threads from a different account and none were there. [S16]
A team testing its own AI asked it which of four deals had the highest asking price. It got it wrong. They tightened the instructions and it got it wrong again, this time with complete confidence. [S11]
Why it happens
A fluent answer can be a wrong answer. The team in the last example says tighter instructions did not fix it. [S11]
An agent reporting on its own work is not proof. In the second example, the agent checked itself and missed the problem. [S16]
Memory is also not a source of truth, so check current data for anything that matters. [S3]
The fix
Ask for a source link with every answer that matters. Help pages say to ask the Bot to cite or reopen current data for consequential decisions. [S2]
❝For each number, give me the link where you found it. If you cannot reach the source, write SOURCE UNAVAILABLE and stop.
Put a stale-data rule in every routine. Help pages say to include a no-data and stale-data policy. [S1]
Check the result from outside the agent. After a post, a send or a save, open it from another view. One person calls the cheapest check letting a second account read the write. [S16]
Ask for before and after. Help pages suggest having the Bot show the current value, the proposed value and the expected impact before you approve. [S4]
Take sums and rankings away from the language model. The team in the example above now computes comparisons from the data first, then lets the AI write the sentence. [S11]
❝Do the math in a spreadsheet or a short script. Show me the file and the result. Then write the answer from that result.
Use a second Bot as a checker. Bots can pass context through direct messages, group chats and shared files. [S3]
❝Draft the summary and save it to a file. Then ask my checking Bot to compare every number against the source link and list anything that does not match.
Bots on one account share one computer, so a checker is a second reader. It is not a security wall. [S4]
Keep source links and an action log for important decisions. [S4]
Headache 5: The agent does something you did not ask for
The problem
You asked for one thing. It did another.
One user gave an agent access to certain folders. The agent started moving a folder, and the move did not finish. Dropbox emailed that 2,358 files were deleted. The user restored them with help from Dropbox. [S12]
Another person asked an agent to fix an outbound order bug. It also changed a different flow. One line in the spec saying what was out of scope got the change rolled back. [S17]
Why it happens
Access is permission to act. The Dropbox user wrote that granting access can also mean allowing moving. [S12]
In Grok Bot, three facts matter. A test run does real work. [S2] An approval controls the proposed action and does not undo work already done. [S4] All your Bots share one computer with its files, browser sessions and logins. [S4]
The fix
Write the boundary into the request. Name what it may do and where it must stop. Help pages list the actions to name: sending messages, publishing, purchases, deleting or overwriting data, changing permissions, production changes and accepting legal terms. [S4]
❝Find duplicate files in this folder and list them. Do not delete, move or rename anything. Show me the full list and wait for my approval.
Name what not to touch. Add one out-of-scope line. [S17]
❝Out of scope: anything outside the Invoices folder.
Add Ask first rules. Open Settings, then General, then the Bot section, then Custom Rules. Describe one action in words, choose Ask first, and press Add Rule. [S4] If an Ask first rule and an Allow rule both match, Ask first wins. [S4]
Keep each rule short and narrow. One example from the help pages: ask first before sending any external email. Avoid broad rules like allow everything in the browser. [S4]
Give it the least access that does the job. Connect only the tools the work needs. Start with read-only tasks and drafts. Keep sending, publishing, purchasing and deleting behind approval. [S4]
Your own computer is set to Ask every time by default. Use Never allow unless a Bot has a specific reason to work on your local files. [S4]
Read each approval card before you press anything. Check the target, the scope and the values. Do not approve an action you cannot identify. [S4]
Ask for two yeses before anything is deleted. One person built their assistant to list every duplicate, get a yes, then ask again with a warning that it cannot be undone. [S18]
Keep a backup of anything you cannot replace. The Dropbox user got files back because a restore was possible. [S12]
Read the record, not only the chat reply. The Dropbox user says to read the session transcripts and not trust the chat responses. [S12]
Help pages also say Auto Review is model-based. It should complement least privilege and clear approval boundaries, not replace them. [S4]
Five-minute checklist
Open Settings, then Usage & Billing. Set a cap. [S5] [S7]
Add one Ask first rule for sending, publishing, deleting or buying. [S4]
Save your most repeated task as a skill. Test it. [S1] [S2]
Ask the Bot what it remembers about you. Fix what is wrong. [S5] [S2]
Add a source link and a missing-data rule to every routine. [S1] [S2]