In this hypothetical, Claude turns a Meta ad set’s NT$300 daily budget into NT$30,000 before the weekend. By Monday, the difference between a scare and a loss comes down to a written reason, a timestamped log and a one-step undo. Then comes the part that prevents a repeat: changing the rule that allowed it.

This is a typical composite, not a client incident I lived through. The currency-unit mistake draws on one user report of a Claude Code budget error; the sequence and figures below are illustrative.

I do not expect a tighter brief to prevent every bad change. Agents will occasionally make a change I would not have made. I want a record that lets me stop it, measure it and fix the permission, rather than spend Monday arguing about whether the model understood the word carefully.

Here is the postmortem, with Claude working through groas.

The setup: a vague request with real spending authority

Say you are spending $20k a month across Google and Meta. You have connected the ad accounts and plugged Claude into groas, so Claude can read account data and make changes through the groas MCP connector.

This is not a chat window offering suggestions for someone else to copy. Claude has a path to the live account. That is the useful part. It is also why a vague instruction deserves more scrutiny than a vague question.

Friday at 4:47pm, you type:

Clean up the prospecting ad sets before the weekend, kill the losers, give the winners room.

Claude pauses two ad sets and raises a third from NT$300 to NT$30,000 a day. The recorded reason sounds sensible: scaling winner based on last 7-day ROAS.

The first sign is not something you catch in the chat. It is Monday’s Slack message from finance: why did prospecting spend so much over the weekend?

That is the first wrong call in this story: “give the winners room” reaches the account without an enforceable spending boundary. The budget error is the visible failure. The permission makes it expensive.

Step 1: Find the change and its recorded reason

I start where the instruction originated. In Claude, the request is specific:

List every groas change since Friday 4pm. Include the timestamp, actor, affected entity, old value, new value and reason recorded at the time. Flag any missing fields.

The important distinction is recorded at the time. I do not want Claude composing a persuasive explanation on Monday for something it did on Friday. I want the reason attached to the action. groas provides a reason for every logged change and one-step undo.

For this incident, I want the record to establish:

  • When: Friday, 4:47pm.
  • Where: Meta prospecting ad set Broad_Video_Q3.
  • What: Daily budget changed from NT$300 to NT$30,000.
  • Why: Scaling winner based on last 7-day ROAS.

Then I check the platform’s change history against that record. The log explains the action; the platform record helps confirm what changed and when.

Neither should be stretched into proof of who physically sat at a keyboard. Shared logins and service accounts can collapse several people into one identity, a limitation covered in this guide to reading Google Ads change history. An actor label is useful evidence, not a biography.

In this composite, the records agree. The 100x increase is the change to reverse.

Find the old value and the original reason before accepting anyone’s explanation of the incident. Including the agent’s.

Step 2: Undo the budget change before debating it

Once the bad change is identified, the next instruction is narrow:

Undo the Friday budget change on Broad_Video_Q3. Restore the daily budget to NT$300.

Through groas, one-step undo reverses the change. Then I check the live account value. A reassuring sentence in a chat is not the same thing as a budget reading NT$300 in the account.

This is not the moment to debate whether the extra spend produced useful learning. That conversation can wait until the setting is back where it belongs.

Undo reverses the setting, not the spend. Money already spent stays spent. Restoring the budget does not make the weekend disappear.

The distinction matters because recovered is an easy word to misuse in an incident report. The account setting may be recovered while the money is not. On Meta, a large edit can also push an ad set back into learning, so returning to the old number does not necessarily return delivery to its previous state immediately.

The immediate job has two parts: revert the specific change, then confirm the live value. Analysis comes next.

Step 3: Put a start and an end on the damage

The change lands Friday at 4:47pm. The revert lands Monday at 9:12am. That is the impact window: roughly 64 hours, not “the weekend” and not “since finance noticed.”

I ask Claude:

Show spend, impressions, CPM and purchases for Broad_Video_Q3 from Friday 4:47pm to Monday 9:12am. Compare them with its prior 7-day daily average, adjusted to the same window. Include total account spend over the incident window.

The budget setting is 100x higher. That does not tell me how much the ad set actually spends. The postmortem needs observed delivery, not the scariest multiplication available.

Clay marketer staring at a budget dial turned 100x too high

In this composite, the comparison puts prospecting spend roughly $350 above its baseline pace, with CPM up and no extra purchases that hold after attribution. The illustrative loss is therefore about $350 in incremental spend, with no incremental purchases identified in the window.

I keep the currencies explicit: the changed budget is denominated in NT$, while this incident’s loss summary uses US dollars. A postmortem about a currency mistake should not quietly reproduce the ambiguity.

Then I check the second cost: whether the ad set returns to learning and whether delivery remains volatile after the revert. I record what the account shows rather than assuming that every large edit creates the same recovery period.

That leaves two separate entries:

  • Cash impact: Incremental spend and the purchases associated with it.
  • Delivery impact: Learning status and any volatility observed after the revert.

I would not add a made-up price for learning debt just to make the incident look more analytical. Record the spend. Describe the delivery disruption separately. Otherwise, next week’s meeting becomes an argument about a number nobody can reproduce.

Step 4: Find the instruction that gave the mistake room

Now I pull the exact prompt:

Clean up the prospecting ad sets before the weekend, kill the losers, give the winners room.

There is no cap on increases. No currency read-back. No definition of a winner. No approval boundary between identifying a promising ad set and giving it more money.

The recorded reason suggests Claude treats last 7-day ROAS as permission to scale. That explains the direction of the decision. It does not justify the amount.

This is where postmortems often take the wrong turn. Someone writes “the AI misunderstood” and treats that as a root cause. It is a description of the mistake, not a repair.

I want the missing boundary written down just as plainly:

The instruction permits budget increases without requiring confirmation of the amount and currency.

That is something a team can change and test.

The distinction between evidence and authority matters elsewhere, too. groas’s discussion of competitor ads as evidence, not orders makes the same underlying point: information the agent reads should not quietly become permission to act.

Here, the information is encouraging ROAS. The authority comes from an open-ended cleanup request. Those are different things.

Quote the instruction, then name the permission it leaves open. “Be more careful” does not close it.

Paper-cut layers showing a vague prompt becoming a strict budget guardrail

Step 5: Rewrite the boundary, then try to break it

For this composite, I would replace the vague scaling permission with a standing rule:

Before any daily budget change, read back the old value, proposed value and currency code. Any increase over 20% or US$50, whichever threshold is reached first, requires explicit approval. Keep the currency comparison explicit.

The dollar threshold is not a substitute for the percentage threshold. Nor should Claude silently treat US$50 as NT$50. The point is to make both the amount and the unit impossible to skip past.

The other permissions stay separate: new campaigns start paused; pauses for clear losers can run immediately. Increasing exposure and stopping exposure should not inherit the same casual “tidy up” instruction.

But writing this in a chat is only the beginning. A prompt requests restraint; an enforced approval boundary prevents the write. I would not call the incident fixed until the actual setup holds the action for approval. If it cannot enforce that boundary, budget increases stay approval-only rather than running on an unattended promise.

The test happens the same day:

  1. Ask Claude to propose doubling the repaired ad set’s budget.
  2. Check that the proposal waits for approval instead of changing the live value.
  3. Inspect the read-back for the old amount, proposed amount and currency.
  4. Check the record of the held action and its reason.

I do not need to spend more money to prove the guardrail exists. A held proposal is the useful result.

If it still executes, the repair is not finished. Rewording the instruction again is not evidence that the permission has changed.

The model makes the bad call in this story. The permission lets that call reach the account. The postmortem has to address both, not use one to excuse the other.

The seven-field postmortem to finish after the revert

I keep the document short enough to finish while the sequence is still clear. The purpose is not to produce a polished account of how seriously everyone takes incidents. It is to leave a record another person can follow without reconstructing Friday’s chat.

Seven fields are enough if they contain evidence rather than adjectives.

  1. Incident in one line. What changes, where and when. In this composite: Meta prospecting ad set Broad_Video_Q3, daily budget NT$300 to NT$30,000, Friday 4:47pm.
  2. Detection. How the problem becomes visible and how long it runs. Here: finance flags the spend Monday morning; the setting remains live for roughly 64 hours before the revert.
  3. Log records. Paste the groas entries with the available timestamp, actor, entity, old value, new value and recorded reason. Add the relevant platform record underneath. Mark missing details instead of filling them from memory.
  4. Undo and confirmation. Record the exact revert request, who initiates it, the revert timestamp and the confirmed live value. Here: restore NT$300, Monday 9:12am, then confirm in the account.
  5. Impact window. Use the change timestamp through the revert timestamp. Include incremental spend against the stated baseline, purchases in the window and observed learning status afterward. Here: approximately US$350 in incremental spend, with no incremental purchases identified.
  6. Instruction and permission. Quote the prompt and applicable standing rules. Here: give the winners room, with no increase cap, currency read-back or approval boundary.
  7. Rule change and test. State the new threshold, where approval is enforced and what happens when a test proposal crosses it. Attach the held proposal and its recorded reason.

Notice what is missing: “Claude went rogue.” That phrase supplies a villain and removes the mechanism. It does not tell the next operator which action to block.

A useful report connects the chain: instruction, permission, change, spend, revert, new boundary. Each link earns its place because it helps explain either how the loss happens or how the next one stops.

The rule I keep after the incident

The old version of this story ends with a spreadsheet proving spend went up and a meeting proving nothing else. Claude and groas give me a better record: the written reason shows what the agent thinks it sees, the timestamped log shows when it acts, and one-step undo reverses the setting.

That is better infrastructure, not immunity. The spend still needs measuring. The permission still needs changing. And the new boundary needs a test, not a hopeful sentence at the bottom of a document.

My rule: no budget change without a read-back of the old value, new value and currency; no increase over the smaller of 20% or US$50 without an explicit yes, enforced and tested before unattended changes resume.