Diagnose

One question: what changed, and when? A diagnosis is a read plus an argument. Reach a cause, name the evidence, then propose exactly one next step — and let marketing-scale or marketing-launch do it.

Prerequisite: the marketing-os foundation skill.

#Method

#1. Pin the claim down

"ROAS dropped" needs three things before it can be diagnosed: which entity, which metric, over which window. Ask if any is missing — guessing produces a confident answer to the wrong question.

#2. Bracket the change

Pull the same metric over the windows the engine actually computes — 3, 5, 7, 15 and 30 days — with from/to (YYYY-MM-DD). The window where the number stops matching the others is the week it happened.

meta_list_campaigns / meta_list_ad_sets / meta_list_ads   with explicit from/to

Compare against the same length of the preceding period, never against a different-sized window.

#3. Decompose before you theorise

A composite metric never moved; one of its inputs did. Walk the chain:

ROAS ↓  →  revenue ↓  or  spend ↑
revenue ↓  →  conversions ↓  or  AOV ↓
conversions ↓  →  clicks ↓  or  conversion rate ↓
clicks ↓  →  impressions ↓  or  CTR ↓
impressions ↓  →  budget ↓, delivery paused, or auction lost (CPM ↑)

Each arrow is answered by a number the read tools already returned. Do not leave the chain until a link actually breaks — that link is the finding, and everything downstream of it is a symptom, not a cause.

#4. Check the known causes, cheapest first

#5. Widen to the other channels before concluding

This is the step a Meta-only tool cannot take, and it is often where the cause actually is:

• meta_get_recommendations — the engine may already have flagged it, with its reasoning chain and the exact metrics it cited. Read it before building your own theory.
• social_list_posts — did organic posting stop? A dead content cadence pulls retargeting pools down a few weeks later.
• social_list_accounts — an expired token means posts have been silently failing. Check before blaming the ads.
• google_search_performance — if organic clicks fell in the same week, the cause is more likely the site or the market than the ad account.
• localfalcon_list_reports — for a local business, a rank slide in the grid and a paid CPA rise are usually the same story.
• get_brand_profile / list_marketing_assets — did the offer or the messaging change? Check the dates against the break.

#Reporting

**What happened**: <metric> went from <x> to <y> between <date> and <date>.
**Where it broke**: <the link in the chain that actually snapped, with numbers>
**Why**: <cause, with the evidence that distinguishes it from the alternatives>
**Ruled out**: <what you checked that did not explain it>
**Next**: <one action>

Ruled out is not padding. It is what separates a diagnosis from a guess, and it is what stops the user chasing the thing you did not check.

#Honesty rules

• If the data does not distinguish two causes, say both. "CPM rose 40% while CTR held — either the auction got more expensive or the audience saturated; frequency at 2.1 argues against saturation, so probably the auction."
• If nothing explains it, say nothing explains it. A confident wrong cause costs the user a week.
• Below 1,000 impressions in the window, say the sample is too small. Small numbers move for no reason.
• null is not zero. ROAS is null at zero spend, CPA is null at zero conversions. "No data" and "zero" are different findings.
• Never quote an industry benchmark as the reason something is underperforming.