kelpi-ai/meta-ads-skills

daily-auditor

The read-only morning audit: last 7 complete days vs the prior 7, wasted budget, fading ads, starved tests, pixel problems, one next test.

查看源码

使用方法

  1. 复制安装命令,在自己的智能体环境安装前先检查源码包。
  2. 从原始文档核对所需工具、凭据和支持的宿主。
  3. 说明输入材料与期望交付物,并对照来源检查输出。
npx skills add https://github.com/kelpi-ai/meta-ads-skills --skill daily-auditor

开始前需了解

  • 安装不包含模型订阅、API 额度或访问文件的权限。
  • 仓库星数和安装量反映活跃情况,不等于独立质量或安全测试。
仓库原始内容

按源仓库内容呈现,保留标题、案例、代码、表格、链接以及原文引用的演示图片。

Daily Auditor

Doctrine

Nobody looks at your ad account every day. Not the agency (check the activity history), not Meta (its recommendations optimize Meta's revenue), usually not even you. Meanwhile the best community advice is "stop touching your campaigns", which only works if something is still watching. The resolution: read daily, act rarely. A read-only audit is what lets you keep your hands off without flying blind.

When to use

  • Every morning, same time, whether things feel fine or not. Quiet failures are the expensive ones.
  • Also: on any account managed by someone else, as an accountability check.
  • Needs a Meta Ads MCP, read access only.

Run it

Run a READ-ONLY audit of my ad account. Change nothing. Compare the last 7 complete days against the prior 7, at campaign, ad set, and ad level, and give me at most 5 findings, ranked by dollars at stake:

1. Wasted budget: spend with zero results, and where it is pooling.
2. Fading ads: flag ONLY if frequency > 4.0 AND CTR is down 30%+ vs the prior window. Name the ad.
3. Starved tests: ads that never got enough spend to be judged (Meta picked a favorite early and starved the rest). Tell me which ads never got a real test.
4. Pixel and conversion health: events that stopped firing, double-fires, or a mismatch that makes the numbers lie.
5. The one next test worth running, based on what the winners share.

For every finding: the numbers behind it, from the account. If a number is not in the account, say "no data", never estimate. If nothing is wrong in a category, say so in one line. End with: "No changes were made."

Guardrails

  • Read-only is the identity of this skill. It proposes; a human disposes.
  • Max 5 findings. A 50-row dashboard dump is not an audit, it is homework.
  • Week-over-week complete days only: partial days and lifetime views both lie.
  • The fatigue rule (frequency > 4.0 AND CTR -30%) is a deliberate two-condition heuristic to avoid false alarms. It is falsifiable; if your data disagrees, tighten or loosen it and write down why.

Good output looks like

Five bullets you can read before coffee, each with real account numbers, ending in "No changes were made."

来自同一仓库

更多 Skills

全部 Skills