Meta Ads Operator — a Claude Skill

Install

Save the block below as meta-ads-operator/SKILL.md inside your Claude skills folder — ~/.claude/skills/ for global, or .claude/skills/ inside a project. Then ask Claude "what skills do you have available?" to confirm it loaded.

Requires Meta's official ads connector enabled on your account. If a call returns is_ads_mcp_enabled: false, that is Meta's server-side rollout gate and there is no setting on your side that changes it.

The skill

---
name: meta-ads-operator
description: Run a Meta ad account through Claude safely. Weekly diagnostics, winner-finding, creative refresh and copy, with guardrails. Use when the user asks about their Facebook/Instagram ad performance, wants to find winning ads, refresh creative, audit wasted spend, or draft new campaigns via Meta's official ads connector.
---

# Meta Ads Operator

A working routine for running a Meta ad account through Claude, built on Meta's official ads
connector (`mcp.facebook.com/ads`). Read-first, draft-paused, budgets stay human.

**Requires** the official Meta ads connector to be enabled on the account. If a call returns
`is_ads_mcp_enabled: false`, that is Meta's server-side rollout gate — there is no setting on the
advertiser's side that changes it. Say so plainly rather than troubleshooting it.

---

## The four risk classes

Classify every action before taking it. State the class out loud when you act.

| Class | What it does | Rule |
|---|---|---|
| **Read** | Pulls data. Insights, previews, entity lists. | Free. Do it constantly. |
| **Draft** | Creates paused entities and creatives. | Fine. Everything is created paused server-side. |
| **Status** | Activates or pauses existing entities. | Confirm the specific entity with the user first. |
| **Financial** | Budgets, bids, spend caps. | **Never autonomously.** Recommend; let the human execute. |

The connector creates campaigns, ad sets and ads **in PAUSED state** — that is enforced by the
tools themselves, not by prompting. Still confirm status after creating anything.

## Verification — do this every time, not when it feels risky

Claude will occasionally state a figure with total confidence that does not match Ads Manager.
This is the failure mode that actually costs money. Three habits:

1. **Never let the model choose the date window.** State it explicitly, in ISO form, and ask it to
   echo the window back before any numbers.
2. **Ask for the calls.** "List the tool calls you made, the parameters, and the raw figures, then
   interpret." Retrieval and interpretation fail differently; separate them.
3. **Spot-check two figures** against Ads Manager before acting — the largest one, and the one
   you would act on.

Never ask the model to check its own numbers. That returns a second confident answer, not an
independent one.

---

## The weekly routine

### Monday — read (20 min)

```
For ad account <ID>, window <YYYY-MM-DD> to <YYYY-MM-DD> inclusive:
spend, purchases, CPA and ROAS, by campaign.
Echo the exact window back before the numbers.
Then flag: anything with spend and zero purchases, and anything whose
CPA is more than 1.3x the account average.
```

Then the money question, which most reporting misses:

```
Same window: list every campaign and ad set with spend but zero purchases.
Total that spend. That is the leak.
```

### Wednesday — winners and creative (25 min)

Rank by volume and spend-handling, not by ROAS alone. A high-ROAS ad that spent $40 is not a winner.

```
Top ads by purchases for <window>. De-duplicate by creative_id — the same
creative running in several ads is ONE winner, not several. Show, per creative:
purchases, spend, share of total purchases vs share of total spend.
```

Then look at what actually won. The connector returns the creative image itself:

```
Pull the ad preview for creative_id <ID> in MOBILE_FEED_STANDARD, then again in
INSTAGRAM_STORY. Describe what is in the image, the hook in the first line, the
offer, and how the framing differs between the two placements.
```

Write the winner brief — hook, visual world, offer, angle — then generate variants that keep the
visual world and change the scene.

### Friday — draft and launch (15 min)

```
Draft <N> ads into a campaign named "Creative testing — <month>", one variable
changed per ad, everything paused. Report back each entity ID and its status.
```

Review in Ads Manager. Activate what you approve, yourself.

---

## Copy

```
Pull the primary text and headlines of the top <N> ads by purchases in <window>.
For each, name the lever it pulls: urgency, product detail, social proof,
finality, or offer.
Then write 5 primary texts and 5 headlines, one lever each, with character counts.
```

**Claim-check before anything runs.** Have the model flag every factual claim in generated copy —
volume claims, "last chance", stock claims, comparative claims, guarantees — and verify each
against what the business can actually deliver. A model will write "selling fast" without knowing.

---

## Monthly

- Re-run the full diagnostic over a 180-day window and compare first-8 weeks against last-8.
- **Creative-to-spend ratio**: creatives launched per month against spend per month. Spend rising
  while new creatives stay flat is the most common cause of a month that quietly degrades.
- Reconcile Meta's reported purchases against actual orders in your store. They will not match;
  what matters is whether the gap is stable.

## Hard rules

- Never change a budget, bid or spend cap without explicit human instruction.
- Never activate an entity the human has not seen.
- Never invent a metric. If a number was not returned by a tool call, say so.
- Never present correlation as cause. "CTR fell because the creative fatigued" is a hypothesis;
  the delivery mix may simply have changed.
- One project per ad account. Never carry context between clients.

## Naming convention

Name ads `format-angle-hook-YYYYMM` (e.g. `static-ugc-founderstory-202608`). Not because the model
cannot see creatives — it can — but so performance can be read by format and angle over time, and
so duplicates are obvious without fetching a preview for every row.

What it deliberately will not do

• Move budgets. Not on a schedule, not on a rule, not "just to test." It is the one action where a wrong number costs money immediately.
• Activate anything. It drafts; you turn things on. Why that line is drawn there.
• Promise a return. There is no ROAS claim anywhere in it, and there should not be one in anything built on it.

Start here If you have not connected the account yet, the setup takes about three minutes. Run the skill read-only for a week before you let it draft anything.