Build a Meta Ads dashboard in Claude

Builds the Get MCP Ads Meta Ads template as a Claude dashboard: KPIs against an equivalent period with results counted per optimisation goal, why CPA and ROAS moved split exactly into CPM, link CTR, conversion rate and order value, a ranked list of what to act on (learning limited ad sets, disapproved ads, delivery issues, fatigued creatives, campaigns held by their budget, the month's landing against budget) with the rule behind each, the campaign, ad set and ad tree with settings, learning stage and targeting, every ad by spend with its thumbnail, review state and rankings, the funnel by campaign, results under each attribution window beside the site's own GA4 count, and the creative, audience, timing and account health views.

Recipe

Intent A live home page for one Meta ad account that says how it is doing, why, and what to do this week, with every campaign, ad set and ad one click away.

Steps: 3
Tools: 1
Rules: 5
Failure modes: 3

skills/meta-ads-dashboard.md

When to use it

Use it when

Someone asks for a Meta dashboard, a Meta account overview, why Meta CPA or ROAS moved, what to fix on Meta this week, or a Meta report they can keep in Claude.
A weekly Meta review before deciding budgets, creatives and ad sets to change.

Do not use it when

Comparing Meta with other platforms: use the Cross-platform scorecard.
Changing budgets, statuses or ads: the dashboard reads only; changes go through the Meta tools with preview and confirmation.
Hosts without Claude's Dashboard artifact type: answer in the conversation with the matching reports instead.

What it needs first

Optional Meta Ads, Google Analytics 4

No single source is required: this one reads across whatever you have connected.

Preconditions

The organisation is on Team or Enterprise: ready-made dashboards are not offered on Free or Pro.
The account is among the chosen accounts, confirmed via list_accounts.

Evidence every conclusion must carry

Every figure carries its account, metric and window.
Anything below a readable sample size is reported as unjudged, not as zero.

Inputs
Input	Required	What it is
`window`		Analysis window.Default: last 30 days.

Procedure

In order. Every tool named is one the gateway ships, and links to its reference.

Steps

Resolve the client's workspace, its Meta ad account and, when connected, its GA4 property with list_accounts.
Fetch the meta template with those accounts (accounts: the Meta account with its currency, and the GA4 property when there is one).get_dashboard_template
Create a Dashboard artifact, write every document exactly as given in one batch, read its check, hand over with the lines given.

Decision rules

The observable condition, what it lets you conclude, and what takes the conclusion back.

When The template is refused with PLAN_REQUIREDConclude Say the dashboards come with Team and Enterprise and that only the account owner changes the plan; answer the question in the conversation meanwhile.
When The person asks why CPA or ROAS movedConclude Read the decomposition: each driver's contribution adds up to the change, so name the largest one and its direction, then look at the campaigns and ads behind it.
When The account mixes objectivesConclude Say results are counted per optimisation goal and never added across types; purchases, CPA and ROAS count the account's purchases over all its spend.
When An alert is questionedConclude Quote its rule from the row and the figures it compared; learning stage, delivery issues and review reasons are Meta's own fields.
When A reader sees a source fail to loadConclude Their own Get MCP Ads access decides: check that their connection covers this client's workspace and that their plan includes dashboards before changing the dashboard.

Where the agent stops

The readout changes nothing. Any change it motivates goes through Safe Writes: preview, human confirmation, then apply.

How it goes wrong quietly

The cases where the analysis is wrong and still looks right. Read them before trusting a number.

Editing the template's page or queries by hand loses the provenance of each figure: ask for changes by name instead.
Adding website, app and all purchases together: Meta counts the same purchase in several families.
Reading a view-through share as fraud or waste: it is how much Meta credits through views, to weigh against the site's own count.