Giving an AI assistant account context: full history or a short brief?

Been thinking about how much context to give an AI before asking it to review a campaign. Some people paste in the whole history: every change log, old ad copy, past tests. Others write a one-page brief with the goal, target CPA, what's already been tried, and what must not change. I think the brief wins most of the time. A full history makes the model anchor on old decisions and bury the current question, and it eats context fast on a big account. The brief also forces you to state the goal, which most people skip anyway. The exception is debugging something that changed over time, like a sudden CPA jump, because then the dated log matters. Curious where others land, especially agencies juggling several accounts.

I'm not sure the anchoring point holds. A brief is only as good as whoever writes it, and what actually causes a CPA jump is usually the thing nobody thinks to mention, like a bid strategy switch or a landing page edit from two months back. Say a dental clinic's CPA doubles and the brief says "nothing changed." The model has no way to know otherwise. Context size is a real limit, but I'd rather filter the log down to the changes that touched that metric than throw the whole history away before anyone knows which part matters.

Agreed on the dated log, but the filter is where people lose the thread. If you only pull changes that touched the metric in question, you can miss whatever changed the auction rather than the account: a conversion action re-pointed, a value rule added, or a site change that broke tag firing. Those often never show up as an edit to the campaign at all. For a dental clinic with a doubled CPA, I'd check first whether conversions dropped or cost rose, because that split tells you which part of the history to open.

Honestly it depends on how many changes landed in the window you're debugging. If the account only had a handful of edits in the last 90 days, just read the whole dated log. It's short enough to scan, and that's how you catch the tag change or value rule nobody thought to mention. If it's dozens of edits a week, the log turns into noise and you won't get anywhere without filtering by date window and metric. So count the changes first, then decide how much to cut.

Counting changes first makes sense to me. I'd go a step further, though: what matters is less the raw edit count and more how many of those edits came from something other than a person clicking around. Scheduled rules, automated bidding shifts, and platform-applied recommendations can change the auction every day without anyone logging a thing, so a "quiet" account can still be the noisiest one. Does anyone here pull the change history and the recommendations feed side by side before deciding how much to cut?