Claude Code audits Meta ads against video scripts

Nicolas Gorroño connects Claude Code to Meta Ads through MCP to investigate losing campaigns for his AI Ranking community. The agent retrieves ad-level spend, purchases and revenue, then writes recommendations with budget-allocation calculations.

It also reads six video transcripts from his project folder and compares creative content with account outcomes. Two losing variants consumed nearly half the budget, while one stronger variant received roughly a fifth. Gorroño subsequently pauses thirteen losing retargeting ads and records new videos from generated concepts. The agent prepares copy for nine new variants in an upload-ready CSV.

The account contains an instructive negative result: those new creatives do not produce purchases. Gorroño retains evidence from the account rather than treating fresh AI-written scripts as guaranteed winners. A scheduled watchdog later pulls recent performance and flags pause or scaling candidates; it is a reporting script, not proof of an autonomous agent changing budgets.

Gorroño reports a sales-campaign return improvement, but the article does not establish controlled attribution to the agent. The useful workflow combines numerical evidence with the actual words used in ads, followed by human recording and campaign decisions. Its recommendations depend on the underlying purchase tracking and on small early samples, so observed creative associations should not be presented as universal causal rules.

Prompt
Suggested prompt (written by Agents In Use, not the author's)

Source: LinkedIn · Nicolas Gorroño