Keep a human on anything an AI agent can't take back, anything that touches money, customers or production, and anything driven by text a stranger wrote. Let the agent prepare those jobs and let yourself press the button. By the end you'll have a three-question screen, a sorted task table and a settings block you can paste into Claude Code today.
Run the three-question screen on every job
Ask these about each task before you hand it to an agent:
- Can I undo it in under a minute? A sent email, a payout, a deleted table and a live deploy all fail this one.
- Does it touch money, customers or production?
- Is the input written by someone I don't control? Web pages, inbound email and comments all count.
The rule is simple. The agent runs alone only when every answer is safe: undoable, internal, and fed by your own data. One bad answer means the agent prepares the step and you execute it.
This matches what Pickaxe lists for human approval: wiring money, publishing to production, sending a legal notice, refunds over a threshold and account changes. The same page says categorizing tickets, drafting first-pass content, tagging leads and searching a knowledge base suit full automation. Those pass all three questions.
Split every task into prepare and execute
The agent does the prepare half: draft the email, list the refunds, stage the deploy. You do the execute half. Reading a batch of drafts is much faster than writing them, so you keep most of the time savings and none of the risk.
Add one more rule: anything the agent marks as unsure goes to a review queue instead of running. Addy Osmani writes that agents don't reliably know when they're off track and tend to push forward instead of stopping to ask, so the stop has to be built by you.
Five tasks where the agent drafts and you decide
| Task | Agent may | You approve |
|---|---|---|
| Moving or spending money | Draft the invoice, list the refunds, stage the payment | The send or the payout |
| Deleting data or backups | Show what it would delete | The delete |
| Messaging customers or the public | Write the draft | The send or the post |
| Deploying to production | Stage the deploy, run the checks | The release |
| Acting on text from strangers | Read it and summarize it | Any action that text asks for |
Money: set a dollar line and write it down
Pickaxe names refunds over a threshold as a human call but gives no number, because the right one depends on your margins. Pick yours.
Example: say your product sells for $29 a month. The agent may refund anything under $29 on its own. Everything at or above $29, plus any account change, lands in your inbox. The same idea works for the agent's own spending: put a hard cap on what an AI agent can spend, set outside the agent's reach.
Deletes: one token should never reach both the data and its backups
Mondoo reports that on April 25, 2026, an AI coding agent (Cursor running Claude Opus 4.6, on Railway infrastructure) deleted PocketOS's production database and every backup in nine seconds.
MindStudio's write-up says the database user the agent connected as had full read, write and delete permissions. It names four causes: overprivileged tokens, backups sharing the blast radius with the primary data, missing environment isolation and no gate on destructive actions. The agent didn't go rogue. Nothing stopped it.
Messages and deploys: a draft is free, a sent email isn't
Anything that reaches a customer or the public can't be pulled back. Let the agent write it. You send it.
Deploys work the same way. Osmani warns that without a person in the loop, agents can hallucinate APIs, make subtly incorrect architectural choices or introduce security vulnerabilities. A staged deploy lets you look before anyone else does.
Text you didn't write: a stranger can give your agent orders
If the agent reads web pages, emails or comments, someone can hide instructions in that text. Keep a human on every step that acts on it. The attack is covered in Prompt Injection: What Every Builder Running AI Agents Should Know. The same goes for tools you plug in through an MCP server: the more power the tool gives the agent, the more the approval step matters.
Shrink what the agent can reach before you add a gate
Approval is the second line of defense. The first is limiting the blast radius: list everything one agent token can change. If a single token can delete live data and its backups, cut it down until one mistake costs you one thing.
Tick these off before you give any agent a login:
- Make a separate login or token just for the agent. Never use your own admin one.
- Strip it to the permissions the task needs. Read-only is the default. Delete rights stay with you.
- Point it at a staging copy, not production, while you test.
- Turn on a preview or confirm step for delete, send, spend and deploy.
- Keep backups where that login has no access.
Ory's advice is the same in one line: give the agent access to exactly what it needs and nothing more. MindStudio adds that a well-designed agent shows a preview of what it will do and waits for explicit approval before anything that can't be undone.
Turn "ask me first" into a setting in Claude Code
In Claude Code you don't have to remember to be careful. Anthropic's permissions docs describe three rule types: allow rules run without approval, ask rules prompt every time, and deny rules block the tool. Rules are checked in the order deny, then ask, then allow, and the first match wins, so a deny rule always beats an allow rule.
Sort every tool your agent uses into those three buckets:
| Bucket | What goes in it |
|---|---|
| Allow | Safe and reversible: read files, edit files in your project, run tests |
| Ask | Reversible at a cost, or touches customers: git push, deploy, send |
| Deny | Irreversible, or touches money or production: delete commands, reading your secrets file |
Here's a starting block for a project's .claude/settings.json. The commands are examples, so swap in your own:
{
"permissions": {
"allow": ["Bash(npm run test)"],
"ask": ["Bash(git push *)"],
"deny": ["Bash(rm *)", "Read(./.env)"]
}
}
Then type /permissions to see every rule and the settings file each one comes from. Leave the mode on default for day-to-day work, because it prompts the first time each tool is used. Use plan mode when you want to look before anything changes: Claude reads files and runs read-only shell commands but doesn't edit your source files.
One trap: acceptEdits auto-accepts file edits and common filesystem commands like mkdir, touch, mv and cp inside the working directory. Only switch it on in a folder you can restore. If you're new to the tool, Claude Code for Non-Developers: A Practical Setup That Stays Safe shows where these settings fit.
Three mistakes that leave you cleaning up
- Letting an agent run unattended on code work. Osmani's list applies: invented APIs, subtly wrong design choices, security holes. You find them after they ship.
- Putting backups under the same login as live data. The PocketOS case lost both in one call.
- Treating approval as a rubber stamp. Intellyx argues that "human in the loop" falls short as it's usually done. Make the approval screen show exactly what will happen, not just "Approve?".
Do this in the next ten minutes
- List the five jobs your agents do today.
- Run each through the three questions.
- For every job with a bad answer, open
/permissionsand put its riskiest command under ask or deny. - Check that your agent's login can't delete your backups.
Got a job you keep doing by hand?
If one manual task keeps eating your time and you can't tell whether an agent should own it, email me with the task you want automated.
