meta-scaling-copilot: scale, hold or cut
What it does
Section titled “What it does”Decides what to scale, hold or cut, and proposes the exact change. Each day it reads trailing stability, learning status, the marginal read (what the next dollar is doing, not the average), net new reach, and creative health. Then it renders one verdict per entity with the specific diff:
- Budget steps capped at 20%, on a 48 to 72 hour cadence, because faster resets learning.
- Cost-cap walks of 5 to 10%.
- Every diff citing a dated decision rule.
Crucially, it grades against your source of record, meaning your store backend or CRM, with the guardrail metric resolved for your declared conversion shape. Never Ads Manager ROAS, which is the number most accounts scale on and the reason they scale into losses.
Read-only in this version. Every change ships as an approved, copy-pasteable change list under a per-diff approval gate. Nothing executes against your account without you saying yes to that specific change.
When to reach for it
Section titled “When to reach for it”- Daily, as the budget check.
- “Can I raise budget on this campaign?”
- Something is working and you want to scale it without breaking it.
- Every account change proposed by the weekly routine lands here for approval.
Where it runs
Section titled “Where it runs”Both.
What it needs from you
Section titled “What it needs from you”- The ad account, confirmed against the platform-reported name before any verdict rests on it.
- Backend revenue access. This is the requirement, not a nice-to-have. Without a source of record the guardrail cannot be computed and the verdicts honestly report that rather than falling back to platform ROAS.
- Your declared conversion shape, which decides which guardrail metric applies.
What you get back
Section titled “What you get back”- One verdict per entity: scale, hold or cut.
- The exact proposed change, sized to the cadence rules.
- The reasoning, citing its dated rule and showing the marginal read.
- A copy-pasteable change list once you approve, diff by diff.
A worked example
Section titled “A worked example”You: “Should I scale campaign 3?”
Verdict: HOLD. Trailing 7-day backend CAC is inside target, but the marginal read is negative: the last 20% of spend is returning at roughly 1.4x against a 2.1x guardrail, so the average is hiding a bad edge. Also 31 hours since the last budget edit, inside the 48-hour cadence floor. Proposed: no change today; re-evaluate Thursday. If it holds: +20%, from $800 to $960/day. Rule cited with its date.
Tips & gotchas
Section titled “Tips & gotchas”- The marginal read is the whole idea. A campaign with a great average and a negative margin is a campaign you should stop feeding, and the average will never tell you that.
- 20% and 48 hours are floors, not suggestions. Bigger or faster steps reset learning and cost you more than the scale gained.
- Per-diff approval means per-diff. There is no approve-all, deliberately.
- No backend access means no confident verdict. It will say so rather than quietly grading you on Ads Manager numbers.
Related skills
Section titled “Related skills”Receives nominations from weekly-account-os, meta-account-doctor and creative-testing-os. Picks up where meta-launch-architect leaves off. Reads the gate verdicts from tracking-doctor and signal-surgeon.