Once bidding, targeting, and placement are automated, over-intervention is the largest controllable cause of underperformance, because every significant edit resets a costly learning phase and most daily swings are noise. This free Claude skill decides whether to intervene, and prices what your tinkering already cost.
I ran the forensics on 90 days of change history and it was brutal, my high-edit weeks were measurably worse and noisier. The noise gate now sits between me and the edit button, and I stopped resetting the learning phase on a single bad day. CPA settled within a month.
Fed it a few weeks of daily conversions and it drew the variance band around my rolling average. Turned out most of what I was reacting to was just day-of-week noise. It also reminded me a target change does not reset learning, so I have a lever that is safe to use.
Best part, it did not say do nothing, it said do less and fix the inputs. The protocol staggers my edits a week apart and the input audit pushed me to value-based bidding and more creative. Discipline over tinkering.
Built for performance marketers running automated campaigns, not data scientists.
Install once in Claude. It prices the cost of your past tinkering from your change history, gates any proposed edit against daily variance so you stop reacting to noise, enforces an intervention protocol that avoids resetting the learning phase, and audits the inputs that actually move an automated account, each with a checks block of honest caveats.
A short setup walkthrough, plus a no-install option if you cannot add skills.
A single prompt you paste into any Claude chat to run the same four modes without installing anything.
Three steps, from your data to a do-nothing or act call.
A change-history export plus daily performance, a few weeks of daily conversions and cost per action, a specific edit you are about to make, or the state of your tracking, bidding, creative, and guardrails.
It tags your significant edits and the learning windows they triggered, computes a rolling average and a variance band sized to your conversion volume, checks a proposed edit against the intervention protocol, or audits your inputs, in code where the math needs it.
The cost of your tinkering, a do-nothing or act call on today's move, a go, wait, or split-and-stagger on the edit, or a prioritized input punch list, each with a checks block flagging thin volume, vendor figures, and what is correlation rather than proof.
The skill works from what you bring, it does not connect to your ad account. It reads your exports, computes, and interprets, and where the change-history analysis is correlational it says so, a prior rather than proof, because seasonality and creative decay are confounders.
In an automated ad account, the honest answer is far less often than instinct says, because every significant edit resets a costly learning phase and most daily performance swings are noise. Here is what resets learning, how to tell a real move from variance, and where your effort actually pays off.
Automated bidding runs on a machine-learning model that must gather data before it can optimize, the learning phase. Meta exits it after roughly 50 optimization events in a rolling 7-day window, and Google Smart Bidding calibrates after about 50 conversion events; during that window delivery is unstable and cost per action is usually worse. A significant edit resets the clock: on Meta the optimization event, audience, creative, or a budget rise over about 20 percent, on Google the bid strategy, conversion actions, major targeting, or a budget change of roughly 20 to 30 percent. The practitioner rule is to keep budget and bid moves under 20 percent per week, and the one exception is a target change, which Google states does not reset learning, so target CPA or target ROAS moves are a low-disruption lever.
Most day-to-day swings in cost per action and ROAS are dominated by statistical noise, small samples, day-of-week patterns, and traffic-mix shifts, not by any change in behavior. The correct lens is a rolling average with a variance band sized to your daily conversion count, noisy daily values around a stable average are variance you leave alone, and only a sustained move outside the band is a trend worth acting on. Reacting to a single great or terrible day optimizes to noise that reverts, the winner's curse and regression to the mean, and the thinner your conversion volume the wider the band and the less you should touch.
The cost is quantified. Meta Marketing Science figures, attributed to Meta through secondary sources and directional rather than independently audited, put a 19 percent lower cost per result on ad sets that exit learning and a 68 percent lower cost per result on accounts that keep under 20 percent of spend in learning versus those above 50 percent. The strongest independent evidence is Optmyzr's three-year Black Friday and Cyber Monday study across tens of billions of impressions, where a manual seasonality override of Smart Bidding drove cost per click about 2 times higher and typically cut ROAS by 10 to 17 percentage points.
When the platform owns bids, targeting, and placement, your leverage moves to the inputs. Conversion-signal quality comes first, the algorithm is only as good as what it measures. Value-based bidding follows, advertisers who moved from target CPA to target ROAS saw a median 14 percent increase in conversion value at similar ROAS in Google's data. Creative volume is the biggest lever on Advantage+ Shopping, up to 150 combinations per campaign, and guardrails such as brand exclusions and existing-customer caps stop the account over-allocating to cheap retargeting. Restraint is not neglect, set and forget also fails, so the discipline is to feed the inputs and verify with holdout tests, not the in-platform dashboard.
The skill runs four modes, FORENSICS to price your past tinkering from your change history, NOISEGATE to tell a real move from variance, PROTOCOL to gate a proposed edit before it resets learning, and INPUTS to audit what actually moves the account. The limits are honest, the change-history analysis is correlational and not proof, platform-attributed figures are directional, and below about 50 conversion events per period the variance math is weak so you should intervene even less. For the full argument and the evidence, from the learning-phase mechanics to the incrementality data, read In an Algorithmic Ad Account, Doing Nothing Usually Wins.