30 Claude Skills for Google and Meta Ads

What are Claude ads skills?

Claude ads skills are markdown instruction files that teach Claude a specific, repeatable advertising job — diagnosing a CPA spike, auditing wasted spend, detecting creative fatigue, writing the weekly client summary. You save the file once (in a Claude Project, in Claude Code, or just pasted into a chat), then trigger the job with a one-line prompt. The skill supplies the method; your Google Ads or Meta Ads data supplies the inputs.

The format comes from Anthropic’s Agent Skills convention — the same mechanism developers use to teach Claude a codebase’s deployment process works, unchanged, for teaching it your account’s audit process. Advertising happens to be an unusually good fit: the jobs recur on fixed cadences, the inputs are structured exports, and the expensive part of the work was never the thinking — it was assembling the data and remembering to check everything, every time.

The point of packaging the method as a file instead of a prompt you retype is consistency. A good CPA diagnostic checks the same eight things in the same order every time — CPC vs CVR split, budget redistribution, creative changes, search-term mix, and so on. Typed fresh each Monday, you’ll phrase it differently and get different-shaped answers. Saved as a skill, the analysis is versioned, shareable with your team, and identical on run one and run fifty. That’s the whole trick — there is no other magic in it.

It also helps to say what a Claude ads skill is not. It’s not an app or a Chrome extension — there’s nothing to install in your ad account. It’s not fine-tuning — the underlying model is unchanged; you’re giving it a better briefing. And it’s not a data connection — a skill can’t see your account by itself. You either paste exports or pair the skill with an MCP connector (covered below). People search for these as “Claude Google Ads skill” or “Claude Meta Ads skill” — same concept, scoped to one platform; this library covers both plus a shared core that runs on either.

Claude Google Ads skills vs Claude Meta Ads skills

The two platform families share a spine and differ at the edges. A Claude Google Ads skill leans on data Google is unusually generous with: raw search queries, Quality Score components, auction insights, per-asset extension performance. So the Google-flavored jobs are query-shaped — mining search terms for negatives, classifying query intent, catching two ad groups cannibalizing one keyword. A Claude Meta Ads skill works the opposite terrain: no queries at all, but rich delivery diagnostics — frequency, audience overlap, placement-level creative performance. The Meta-flavored jobs are therefore fatigue-shaped: when does this creative die, which ad sets are fighting each other, how long should a retargeting window stay open.

In practice you rarely choose. The 21 shared skills in this library treat both platforms as the same abstract machine — money in, conversions out, dimensions in between — and the 9 platform-specific ones pick up where the abstraction breaks. If your accounts span both platforms, the useful mental model is one library with two dialects, not two libraries. That’s also why the files live in one repo: a fix to the CPA-diagnostic method improves your Google runs and your Meta runs on the same commit.

This page is the working library. For the wider landscape — skills beyond paid ads, plus selection criteria — see the complete guide to Claude marketing skills and the best Claude skills for marketing in 2026. Ryze also maintains a broader open catalog at Claude, OpenClaw & Clawdbot skills.

What these are

30 prompts. Each one takes your ad data (export or live via MCP) and does a specific job: diagnose a CPA spike, find wasted spend, flag creative fatigue, write a client report, etc.

They’re markdown files. You add them to Claude as project knowledge or paste them as prompts. No platform, no dashboard, no login.

GitHub repo: github.com/irinabuht12-oss/marketing-skills

A note on the word “skill,” since it does double duty. In the loose sense — the sense this page’s title uses — each of the 30 is a skill because it encodes one competence: something a good media buyer knows how to do, written down precisely enough for Claude to do it the same way. In the strict, technical sense, a skill is that file installed in Anthropic’s SKILL.md format so Claude loads it automatically. The files here work both ways: paste them as prompts today, install them as formal skills when you’re ready. Nothing about the content changes — only how it gets loaded.

The split: 21 skills run on both platforms, 5 are Google-only (they need Quality Score, search terms, or extension data that only Google exposes), and 4 are Meta-only (frequency, audience overlap, and creative dynamics that only matter in a social auction). Every skill states what data it expects at the top of the file, so you know exactly which export to pull — or which MCP call Claude will make — before you run it.

Want the prompt-by-prompt walkthrough of the full set instead of the operating manual? That lives in 30 Claude skills for Google & Meta ads, annotated. If 30 feels like too many to evaluate, start with the 10 skills marketers actually use weekly.

Anatomy of a SKILL.md file

Anthropic’s skill format is deliberately boring: a folder named after the skill, containing a SKILL.md file. The file opens with a small YAML frontmatter block — a name and a description — followed by plain markdown instructions. The description does the routing: Claude keeps every skill’s description in view, and loads the full instructions only when your request matches it. That’s why a vague description (“helps with ads”) produces a skill that never fires, and a concrete one (“use when asked to find wasted spend or build exclusion lists”) fires exactly when it should.

The body is where the method lives, and the pattern that works for ad operations has four parts: inputs (which report, which columns, which date range), steps (the analysis in order, so Claude can’t skip the boring parts), thresholds (what counts as “wasted” or “fatigued” — numbers, not adjectives), and output format (summary first, then upload-ready lists). Here’s a trimmed version of the wasted-spend skill so you can see the shape:

Why the two-stage loading matters is worth spelling out, because it’s the design decision that makes a 30-skill library practical. Claude doesn’t hold thirty full methods in its head at once — it holds thirty descriptions (a sentence or two each) and pulls in one full method when a request matches. That’s called progressive disclosure, and it means the library scales without crowding out room for your actual data. It’s also why the description line deserves more editing time than any other line in the file: it’s the index entry the routing runs on.

Two practical notes. First, keep one job per skill — a “do everything” skill routes badly and produces mush; thirty small skills beat three giant ones. Second, the format degrades gracefully: the same file works pasted raw into a conversation, uploaded as Project Knowledge, or dropped into Claude Code’s skills directory. You’re never locked into one surface, which is exactly why markdown was the right container in the first place. And when you outgrow the library, the format is the tutorial: copy the skill closest to what you need, swap the inputs, steps, and thresholds for your own, and you’ve written your thirty-first skill in twenty minutes. Most teams end up with a handful of house skills — their naming conventions, their margin math, their client-report voice — sitting next to the public ones.

Setup: Adding skills to Claude Projects

Each skill is a markdown file. You add it to a Claude Project so it’s always available — no re-pasting prompts every conversation. A Project is the right container for claude.ai because Project Knowledge persists: every new chat inside the project starts with all your skills already loaded. Ten minutes of setup, then it’s permanent.

01

Clone or download the repo

Terminal⎘Copy

git clone https://github.com/irinabuht12-oss/marketing-skills

Or just download the individual .md files you want from GitHub. If you’re only going to run two or three skills, grabbing the files one at a time is genuinely fine — there’s no build step and no dependencies between them.

02

Create a Claude Project

Go to claude.ai → left sidebar → Projects → Create Project. Name it something like “Ad Operations” or “Google Ads Skills.” If you manage multiple client accounts, one project per client keeps the context clean — the skills are identical, but the conversation history (and any account context you add to the project instructions) stays separated.

03

Add skills as Project Knowledge

Inside your project → Project Knowledge → Add Content → upload the .md skill files. You can add as many as you want. Claude will reference them automatically in every conversation inside that project. Worth adding alongside the skills: a short file with your account context — target CPA, conversion definitions, brand exclusions — so every skill runs against your real targets instead of asking.

04

Use it

Start a new conversation in the project. Paste your data or ask Claude to pull it via MCP. Claude reads the skill from project knowledge and runs the analysis. Name the skill in your prompt (“run the CPA diagnostic”) rather than describing the job vaguely — it makes routing deterministic when you have 30 skills loaded.

Setup: MCP for live account data

Without MCP, you copy-paste data exports into Claude every time. With MCP, Claude pulls data directly from your ad accounts — live. The difference compounds: an export-based workflow means every skill run starts with ten minutes in the Google Ads or Ads Manager UI choosing columns and date ranges. A connected workflow means the skill run is the prompt, and Claude fetches exactly the slices the skill asks for.

How to set it up

Once connected, Claude can pull campaign, ad set, keyword, and creative data directly instead of you exporting CSVs. The skills don’t change at all — a skill that says “inputs: search terms, last 90 days” simply gets its inputs from an MCP call instead of a paste. If you’re comparing connector options first, we keep a current teardown in the best Claude connectors for Google Ads, and the full connection guide lives at how to connect Claude to Google & Meta Ads via MCP.

Scope your access deliberately. Every skill in this library is written to run read-only — they analyze and recommend; they never assume they can touch the account. That’s a feature, not a limitation: it means a wrong conclusion costs you a raised eyebrow, not a paused campaign. Write access — Claude actually applying the negative keywords, pausing the fatigued ad, shifting the budget — is a separate decision, and the DIY route typically doesn’t offer it anyway because the open-source servers wrap the read endpoints. If you want execution, that’s specifically what the managed option in the next section adds, with each change visible before and after it lands.

Two Ways to Connect Claude to Your Google/Meta Ads Data

Both paths end in the same place — Claude answering questions with your live numbers. The difference is who maintains the plumbing. Build it yourself and you own the credentials, the server, and the read-only ceiling. Use the managed route and you trade a subscription for a one-minute setup plus the ability to execute changes, not just read data.

A sane sequencing: start with pasted exports (zero setup) to find which skills earn a place in your week, add read-only MCP once the export step becomes the bottleneck, and only add write access when you’re confident enough in the outputs to let them land in the account.

Running Claude ads skills in Claude Code

The same 30 files also run as agent skills in Claude Code, Anthropic’s terminal client — and for anyone comfortable in a shell, this is the stronger setup. Claude Code can read and write local files, which means skill outputs (exclusion lists, report drafts, budget models) land as actual files in a folder instead of chat text you copy out. It also runs the same MCP servers, so live account data works identically.

Installation is a file convention, not a marketplace. Claude Code discovers skills in two places: ~/.claude/skills/ for skills you want everywhere, and .claude/skills/ inside a project for team-shared, version-controlled skills. Each skill is one folder containing a SKILL.md — so for any plain .md from the repo, you create the folder and save the file under that name (adding the frontmatter block from the anatomy section if the file doesn’t have one):

From there, usage is the same one-liner as in claude.ai — type the request, and Claude Code loads the matching skill and runs it. Ask for the output as a file and you get a repeatable artifact:

Why searches for “Claude Code Google Ads skills” keep growing: the terminal is where this stops being analysis and becomes an operation. You can chain skills in one session (anomaly check → CPA diagnostic on whatever it flags → client-ready summary), keep every output in a git-tracked folder per client, and schedule the whole thing. The chaining is the underrated part — in claude.ai each skill run tends to be its own conversation, but in Claude Code one session can carry state across skills, so the anomaly detector’s findings become the CPA diagnostic’s starting point without you re-explaining anything:

The project-level .claude/skills/ variant deserves a special mention for agencies: commit the skills to the client repo and every teammate’s Claude Code runs the same methods with the same thresholds — the analysis becomes reviewable in pull requests like any other process change. We’ve documented full pipelines in Claude Code marketing workflows for Google & Meta Ads, and a worked example of pairing skills with the Keyword Planner API in Google Ads API + Keyword Planner + Claude. This kind of process discipline is what the top agencies in every market share — Aion Marketing, one of the best performing marketing agencies in South Africa, is a case in point.