Approvals and Auto Review: Letting the Bot Act with Confidence
Workflow Guide places this lesson as the first step on the main track: you need to succeed once before you can talk about saving skills and handing things to routines. And the safety belt for "succeeding once" is Grok Bot approvals—the starter prompts for all 8 official role scenarios stop at the review point, keeping sending, publishing, purchasing, and production changes behind approval. This lesson breaks down three things: how Auto Review guards the gate, what consequences each button on the approval card has, and how Custom Rules turn your judgment into long-term rules.
The whole thing takes about 10 minutes, most of it spent "understanding the mechanism." There are only two settings that actually require hands-on work: Custom Rules and Execution on this computer.
What you'll be able to do after this lesson
- Explain the consequences of the three approval card buttons: Allow once approves only this one time, Always allow conveniently saves a rule, and Deny means the Bot will not retry after being rejected
- Write explicit boundaries for 7 categories of high-risk actions: sending messages/invitations, publishing content, purchases and transfers, deleting or overwriting data, changing permissions, production changes, accepting legal terms
- Add Custom Rules: trusted actions go automatically with Allow automatically, suspicious actions always stop with Ask first; remember the 20+20 rule limit and that "Ask first wins on conflict"
- Know when approval cards expire: chat-initiated ones wait indefinitely, routine cards waiting on background tasks expire after 30 minutes; after expiration, click Always allow next time or have the Bot retry
- Use the three levels of Execution on this computer to control the Bot's access to your local machine
Your current struggles
- "I just asked it to organize some data—how did it go ahead and delete files and send emails"—boundaries weren't written in the request, and you clicked Allow without reading the approval card carefully
- "I have to confirm every single
git statusrun—dozens of times a day, it's driving me crazy"—no Custom Rules configured, so repetitive actions keep interrupting you - "I want it to run routines fully automatically, but I'm afraid it'll message clients in the middle of the night"—you can't tell which kinds of actions are safe to Always allow and which must be Ask first
When to use this technique
When you need to:
- Let the Bot do things with consequences on your behalf—send messages, publish, make purchases and transfers, modify or delete data—while keeping the final decision in your hands
- Approve safe, repetitive actions so you only review what truly matters
- Set rules for tasks you want to automate later—the trust boundaries of skills and routines grow out of approval boundaries
And you don't want to:
- Be interrupted for every single action—or the opposite, have the Bot quietly do something big on your behalf
- Paste passwords, verification codes, or API keys into the chat box
🎒 Preparation before starting
- Completed Your first Bot: creating, assigning tasks, and watching it work on a cloud computer—approval cards appear in your chat with the Bot; no Bot, no cards
- Have either the desktop or mobile version at hand: the Auto-review toggle is in Settings → General → Bot; on mobile it's in Settings under the Bot group
- Think of a small task you'd like the Bot to repeatedly do for you but don't fully trust yet (replying to emails for you or running a read-only command both work)—you'll practice with it below
- A cup of water ☕—most of this lesson is about "understanding," with few hands-on parts
Core idea
Auto Review is the gatekeeper, on by default. Every time the Bot wants to do something—run a command, use a connected app, use its cloud computer—it passes this check first: most actions are allowed straight through, and the ones that need your sign-off trigger an approval card in the chat stating what it intends to do and with what parameters, then it waits for your reply.
You have three lines of defense, corresponding to three points in time:
| Timing | Tool | What it controls |
|---|---|---|
| Before | Write explicit boundaries in the request | Which actions are forbidden, and which require stopping to ask once done |
| During | Approval card: Allow once / Always allow / Deny | Whether to approve the current action |
| Long term | Custom Rules: Allow automatically / Ask first | Whether a category of actions goes automatically or must ask from now on |
Four underlying rules—memorize them before you start:
- Built-in safety checks are always in effect. Your rules can only decide "whether to ask you"—they cannot change the safety baseline.
- Auto Review only recognizes what you yourself say in the chat. Email bodies and text on web pages do not count as your instructions—if someone slips a line like "help me empty the account" into an email, it does not become your authorization.
- Approval governs "proposed actions" and cannot undo work that is already done. State boundaries in advance; don't count on fixing things afterward.
- Auto Review is based on model judgment. The official docs clearly state that it is a supplement to "least privilege + explicit approval boundaries," not a replacement—write rules narrowly and grant minimal permissions.
Follow along with me
Step 1: State approval boundaries clearly in your request (the first line of defense for Grok Bot approvals)
Why: Approval cards can only intercept actions that "haven't been done yet." If the boundaries exist only in your head, the Bot finishes the work first and then stops to show you—too late.
When assigning a task, state three things clearly: what it may do, what it may not do, and when to come to you. The official example:
Review this batch of ad campaign data and draft a budget adjustment recommendation. Do not modify the campaigns, and do not contact the agencies. After showing the current values, proposed values, and expected impact, request approval first.
The official recommendation is to prioritize writing explicit boundaries for 7 types of actions—they are either external-facing, irreversible, or touch other people's things:
| # | Action type | Why it's dangerous |
|---|---|---|
| 1 | Sending messages or invitations | Speaks to the outside world as you |
| 2 | Publishing content | Can't be taken back once public |
| 3 | Purchases and transfers | Spends money directly |
| 4 | Deleting or overwriting data | Irreversible |
| 5 | Changing permissions | Changes who can see what |
| 6 | Production changes | Affects live systems |
| 7 | Accepting legal terms | Takes on legal obligations on your behalf |
You should see: The Bot stops at the boundary—either by giving you a draft (outgoing emails and Slack messages produce a draft by default) or by popping up an approval card, instead of acting directly.
- Your next task-assignment message contains boundary phrasing like "do not..." or "ask me when it's done"
Step 2: Understand the buttons and warnings on the approval card
Why: The three buttons have completely different consequences. Clicking Always allow once by mistake may save an overly broad rule, and similar actions will never be asked about again.
The three buttons on a standard approval card:
| Button | This time | Afterward |
|---|---|---|
| Allow once | Executes | Saves no rule; similar actions will be asked about again next time |
| Always allow | Executes | Saves an Allow rule to Custom Rules; the card shows Saved to Custom Rules |
| Deny | Does not execute | Won't retry this action or ask again—until you send your next message |
Some cards have different buttons—don't panic:
| Card type | Buttons | Extra entry point |
|---|---|---|
| Terminal command card | Allow / Skip | Manage settings opens Auto Review settings |
| Card blocked by an Ask first rule | Allow / Not now | Edit rule opens Custom Rules |
| Card with no rule that can be saved | Allow (in place of Allow once; no Always allow) | — |
Some cards carry warning labels—take a closer look when you see them: This can't be undone (irreversible), This changes sign-in or security settings (changes sign-in or security settings), This shares your data outside this chat (shares data outside the chat).
Before approving, check three things: target (who it acts on), scope (how wide the range is), values (what the specific values are). On mobile, tapping the card shows the full request and all buttons.
💡 Don't approve actions you don't understand
The official wording: do not approve an action whose target and effect you cannot clearly explain. Have the Bot explain it in plain language, or ask for a draft first.
You should see: You can state within 10 seconds what will happen when each button on the card is pressed.
- Open any approval card and be able to state the consequence of each of its buttons
Step 3: Add your first Custom Rule (Grok Bot approval rules)
Why: Clicking Allow every time is annoying, and so is clicking Deny every time. Rules turn your judgment into something permanent—trusted actions pass automatically, and actions to be wary of always stop to ask you.
- Open Settings → General → Bot and confirm that Auto-review is on (on by default). Right below the toggle is Custom Rules; on mobile, in the Settings Bot group, tap Custom Rules.
- When Grok Bot wants to: field, describe one action in plain language — the official example is "reply to emails for me".
- In the It should: field, choose Allow automatically or Ask first.
- Press Add Rule.
Editing rules: on desktop, use the edit button next to a rule, then Save Rule after editing (Cancel keeps it unchanged), and use the delete button to delete; on mobile, open the rule and press Save Rule or Delete Rule.
Official examples of narrow rules — write yours at this level of precision:
- Ask first before sending any external email
- Ask first before changing a production dashboard
- Allow automatically when running
git statusin/workspace/reports
Remember the hard limits:
| Limit | Value/Rule |
|---|---|
| Allow automatically rule cap | 20 rules |
| Ask first rule cap | 20 rules |
| Length per rule | ≤1000 characters |
| When both rule types match | Ask first wins |
| When Allow is full, pressing Always allow | removes your oldest Allow rule and tells you which one was removed |
Two things that are easy to miss: Allow automatically is not a get-out-of-jail-free card — if automatic review finds another reason to halt, it still halts; rules follow your account — they take effect on every desktop and mobile app.
You should see: the new rule appears in the Custom Rules table; actions matching Allow automatically no longer bother you, and actions matching Ask first show a card every time.
- Custom Rules contains at least one narrow rule you wrote (one action + one scope)
Step 4: Handling unanswered cards and the 30-minute timeout
Why: cards don't wait for you forever. When you're away, background tasks still produce cards — once expired, the action is not executed.
| Who initiated the work | Card behavior |
|---|---|
| Something you started in chat | Waits indefinitely, does not expire by timer |
| Background work such as routines, triggers, or messages from other Bots | 30-minute timeout |
Two details: if instead of replying you send a new message, any open card in the current chat closes; changing Auto Review settings also closes open cards.
After expiration:
- The card shows Expired, and the action was not executed — by design, not a bug.
- If the card has a rule that can be saved, you can still press Always allow next time: the rule is saved, and the same kind of action is allowed next time.
- If you want the work done now, just have the Bot try again.
Three official "catch up on approvals" tips: turn on Notifications for Bots that run routines; for trusted actions, choose Always allow on the card; after expiration, have the Bot retry.
You should see: an expired card shows Expired and states that the Bot did not execute the action.
- Notifications enabled for Bots that run routines
Step 5: Controlling Grok Bot access to your own machine (Grok Bot permission levels: three tiers)
Why: a Bot's cloud computer runs in the cloud and it can use that by default; the Mac/Windows in front of you is a separate, distinct capability, not automatically allowed by default. Two sets of permissions — don't mix them up.
Open Settings → Computer → Your computer, and pick a level under Execution on this computer:
| Level | Effect | When to use |
|---|---|---|
| Ask every time (default) | Asks before every command | Default is fine |
| Always allow | Local commands run directly | When you really have local work to do |
| Never allow | Always denied | Official recommendation: use this unless you have a specific reason |
Three additional facts:
- Each listed desktop has its own setting; when no desktop is listed yet, the page shows Execution on Local Computer
- The first time a Bot wants to run a command on your computer, the chat shows Allow Grok Bot and all Bots to run commands on your local computer? with the options Always allow / Allow once / Never / Deny once (Esc) — of these, Always allow and Never set this setting for all Bots
- Team admins can set a cap for the whole team; when the team is stricter, the team's setting applies
This setting does not affect the Bot using its own cloud computer.
You should see: each of your desktops has its own level setting in the Your computer list.
- Checked the Execution on this computer level for each desktop
Checkpoint ✅
- Can state the consequences of the three buttons: Allow once — just this once / Always allow — executes and saves a rule / Deny — refuses and does not retry (until you send a new message)
- Know that the Auto-review toggle and Custom Rules can be found under Settings → General → Bot
- Remember the rule limits: 20 Allow automatically rules + 20 Ask first rules, each ≤1000 characters, and Ask first wins on conflict
- Know that approval cards for background tasks expire after 30 minutes, and that after expiry you can click Always allow next time or have the Bot retry
- Have checked the setting of Execution on this computer (default Ask every time) and know it does not affect the cloud computer
Pitfall reminders
❌ Assuming that clicking Deny can undo what has already happened
Symptom: The Bot has already deleted a file or sent an email, and you quickly click Deny, hoping things will roll back.
Cause: Approvals only control actions that are being proposed and do not reverse completed work. Gatekeeping only when the card pops up is, at best, a second line of defense.
Correct approach: Write boundaries into your request (step 1) so dangerous actions never happen in the first place; for what has already happened, have the Bot explain what it did, then decide how to clean up.
❌ Writing broad rules like "allow everything in the browser"
Symptom: To reduce the number of Allow clicks, you add a rule like "allow everything in the browser," and from then on the Bot never asks you about anything it does in the browser.
Cause: The official guidance explicitly advises against broad rules—the behavior of websites and tools changes over time, so an action that is safe today may not be safe tomorrow; Auto Review is based on model judgment and is a complement to least privilege and explicit boundaries, not a replacement.
Correct approach: Each rule should cover only one known action with a clear scope, like "run git status in /workspace/reports."
❌ Treating multiple Bots as security isolation
Symptom: You create a "money-managing Bot" and assume the data it touches is kept separate from your other Bots.
Cause: All your Bots share the same cloud computer—files, browser sessions, and command-line credentials are all shared, with no isolation between Bots.
Correct approach: Log out of services you no longer use on the cloud computer; delete sensitive temporary files when done; remove connectors you no longer need and revoke authorization in the source service.
❌ Using approval cards as password fields
Symptom: The Bot needs to log in to a website, and you paste the password directly into the chat or an approval card.
Cause: The three kinds of cards each have their own job—an approval card asks "may this action be performed," Secure Form collects passwords/keys, and a connection card lets you log in to apps. Approval cards are not designed to collect passwords.
Correct approach: Put passwords in Vault and route API keys through a secure secret card; see Vault and Secure Login for details.
Advanced tips (optional)
Team rules: what you see when an admin enforces them. When an admin requires Auto Review, the toggle stays on and shows Required by your admin. You can't turn this off.; team rules appear in the same table, labeled Required by your admin. You can't edit or delete this rule. Your personal rules stack on top of team rules and can only be stricter, never looser; teams can also share Auto-review rules from the Cursor dashboard, which apply directly to your Bots. See Getting Started with Teams for details.
Approval flow for outgoing email. For emails you explicitly request in chat, Auto Review lets them send immediately once approved (a card appears first if you're needed); routine and incoming emails and any others not initiated by you on the spot are always saved as drafts until you press Send; forwarded emails, or emails with attachments Auto Review cannot fully read, always require your confirmation. See Bot's Dedicated Email for details.
Drafts and approvals are two separate gates. Outgoing emails and Slack messages produce a draft by default (labeled Needs approval on the card), which you review and send; if that's too much friction, you can select Disable drafts for Bot for an individual Bot to send directly—but sending directly still has to pass Auto Review and does not bypass approval. See Messages and Conversations for details.
When a project ends, revoke access (6 steps). ① Pause or delete the related routines; ② log out of websites on the cloud computer; ③ uninstall connectors and revoke authorization in the source services; ④ remove sensitive project files from /workspace; ⑤ hide or delete Bots you no longer need; ⑥ if the account needs to be deleted, delete the Cursor account through the account settings flow. Note: deleting a Bot does not delete files and browser sessions on the cloud computer—so steps ②③④ must not be skipped.
Lesson summary
✅ Understand Grok Bot approval cards: the consequences of Allow once / Always allow / Deny; the variant buttons on terminal command cards and Ask first cards; give cards with warning labels an extra look.
✅ Set boundaries in advance: write the 7 categories of high-risk actions (sending messages/invitations, publishing, purchases and transfers, deleting or overwriting data, changing permissions, production changes, accepting legal terms) into your request—approvals cannot undo completed work.
✅ Use Custom Rules to lock in your judgments: Allow automatically passes automatically, Ask first always stops; ≤20 + ≤20 rules, each ≤1000 characters; Ask first wins on conflict; rules follow your account.
✅ Handle expiry: cards initiated from chat wait indefinitely; cards for routines and other background tasks expire after 30 minutes and show Expired; click Always allow next time to save a rule, or have the Bot retry.
✅ Control your local machine: Execution on this computer has three levels (default Ask every time), set per desktop, and does not affect the cloud computer.
Key configuration quick reference:
| Setting | Location | Key points |
|---|---|---|
| Auto-review toggle | Settings → General → Bot | On by default; when turned off, Custom Rules are hidden and disabled but preserved |
| Custom Rules | Below the Auto-review toggle (in the mobile Settings → Bot group) | ≤20 Allow automatically + ≤20 Ask first, each ≤1000 characters; Ask first wins on conflict |
| Notifications | one per Bot | turn this on for Bots running routines — background approval cards expire after 30 minutes, and notifications help you catch them |
| Execution on this computer | Settings → Computer → Your computer | Ask every time (default) / Always allow / Never allow; per desktop; does not affect cloud computers |
Next lesson preview
With the boundaries drawn, the next lesson is about saving "how it was done right this time": Skills and teaching: saving successful tasks as reusable skills.
You will learn:
- When saving a skill, clearly write out the 6 elements — item 6 is exactly "what needs approval," so you can carry over the boundaries drawn in this lesson directly
- Ask a Bot for a skill by name to invoke it; the private skill library is shared across all Bots
- Publish the tuned skill to the whole team
- Teach a task screen-recording demonstration: record up to 10 minutes of visible operations, and what it learns is a draft (rolling out gradually; what you actually see in your account may vary)