Standing brief for AI ad work: save it once or paste context each session?

Once an AI assistant is connected to your ad accounts, you have to decide where the context lives. One option is a standing brief saved in a project or custom instructions, so it always knows your target CPA, which brand terms you can't bid on, and that budget changes need your sign-off. The other is pasting the context in at the start of each session. Most people end up doing one or the other without really thinking about it.

The saved version is convenient, but it goes stale. Targets change, a product gets discontinued, and the old rules keep steering recommendations long after they stopped being true. So my view is to save the stable stuff, like margins, brand terms and what counts as a lead, and paste the volatile stuff, like this week's budget or the current test. A stale rule is harder to spot than a missing one.

A stale rule is dangerous because it reads like a fact, so the assistant just applies it without hesitation. One cheap fix is putting a "last confirmed" date next to each saved rule, like "Target CPA $40, confirmed Feb 3." That gives the assistant something to check, and when a recommendation leans on an old date it can flag that the rule may be out of date instead of quietly steering by it.

Who usually owns updating those dates, you or whoever runs the account day to day?

I don't think the assistant will catch a stale date on its own. It doesn't reliably know today's date, and it can't tell whether a CPA target still holds. A "confirmed Feb 3" line only helps if you tell it what to do with that line. So put the rule in the instructions: "treat anything confirmed more than 60 days ago as provisional, and ask before relying on it." That way the check lives in the instructions instead of depending on the assistant's guesswork.

For ownership, I'd give it to whoever approves budget changes, not the person running the account day to day. The approver is usually the one who hears when margins shift or an offer gets pulled. The day-to-day person tends to inherit the old number and keep using it because nobody told them it changed.

The approver makes sense to me. In e-commerce, though, the number that goes stale fastest is usually the CPA target itself, because it's derived from margin. A flat "target CPA $40" hides where it came from, so if unit cost rises or the shipping subsidy goes up, the target is wrong and nobody sees the arithmetic change.

I'd save the inputs instead of the result: product cost, average shipping subsidy, payment fees, and typical discount rate. Put a "confirmed" date on each one. When a supplier email arrives, the approver knows which line to update, and the assistant can recompute the CPA from those inputs rather than quote an old figure. A store running a 20% sitewide sale, for example, should have that discount in the inputs, so the target shifts for the week it runs instead of being steered by the full-price margin.

It's more work up front, but it turns a stale rule into one with a visible source.

Saving the inputs instead of the target is the right move, and it fixes what I was circling around: a rule with no visible source can't be audited. I'd also store each input with its units and the date it applies from, not just a bare value. "Shipping subsidy $6" means something different per order than per item, and a sale discount is only accurate if the stored number says whether it's a sitewide percentage or a margin-after-discount figure. Otherwise the assistant recomputes correctly and the answer is still wrong.

Units matter, but I'm not sure the example holds. A sitewide discount rarely moves the CPA target the way it suggests. If you're bidding on a CPA goal, the target should follow your contribution margin per order, and a 20% sale mostly hits your ROAS floor, not the CPA. I'd also log the sale as its own dated input, so you can tell which one actually moved.

I'm not sure that holds for sale periods. Yes, the CPA target should follow contribution margin per order, but a 20% sitewide discount comes straight out of that margin. Take a $100 order with $40 of contribution at full price. At $80, with the same product and shipping costs, you're left with about $20, so the break-even CPA roughly halves. The ROAS floor moves too, but the CPA ceiling is just margin in cash terms, so it drops a lot. Logging the sale as a dated input is a good idea, but it's also the reason to recompute the target during the sale instead of assuming it stays put.