Scalewell / Blog

Claude skills for marketing: what a paid media team should build first

A Claude skill is an onboarding guide Claude reads on demand; here is what one is, where it runs, and the six skills a paid media team should build in order, from the weekly readout to the account audit.

Labeled folders feeding into a laptop connected by a cable to a performance dashboard, representing Claude skills for paid media.

Marketing skills for Claude are folders of instructions that Claude loads only when a request matches them, so a paid media team can write down how it does a weekly readout once and have Claude follow that method every time. A skill is a SKILL.md file with a name, a description of when to use it, and the steps.

You have probably seen the lists already: fifty marketing skills to install, ninety-eight free skills, a bundle for every channel. You install a pack, ask Claude to audit the account, and get back a fluent summary that reads well and tells you nothing you could act on, because the skill never saw your account. That gap between a skill that describes a method and a skill that runs on your data is where most paid media teams get stuck.

This article is for the head of paid media or the PPC lead who runs the accounts, so it treats Claude skills for marketers as an operations question rather than a feature list. It explains what a Claude skill is as of November 2026, from Anthropic's documentation, then gives an ordered list of the skills a paid team should build first and the structure of each one.

A Claude skill is an onboarding guide Claude reads on demand

A Claude skill is a directory with a SKILL.md file at its root, and that file starts with YAML frontmatter holding two required fields: name and description. Anthropic's Agent Skills documentation describes skills as "reusable, filesystem-based resources that give Claude domain-specific expertise: workflows, context, and best practices that turn a general-purpose agent into a specialist." Anthropic's engineering team put it more plainly in its post of October 16, 2025: "Building a skill for an agent is like putting together an onboarding guide for a new hire."

The feature name has stayed stable. Anthropic launched Agent Skills on October 16, 2025 across the Claude apps, Claude Code and the API, and on December 18, 2025 published the format as an open standard at agentskills.io, so the same folder can run in other agents that adopt it. As of November 2026 the documentation marks the feature as stable, and Google ships its own skills in the same format.

What makes skills workable for a busy team is how little one costs while it sits unused. Anthropic calls the design progressive disclosure and gives the numbers: only the name and description load at startup, at roughly 100 tokens per skill, the SKILL.md body loads when the skill triggers and should stay under 5,000 tokens, and bundled reference files or scripts load only when Claude reads or runs them. So you can keep a dozen paid media skills installed and pay for one at a time.

Hard limits apply to the frontmatter. A name is at most 64 characters, lowercase letters, numbers and hyphens only, and cannot contain the words anthropic or claude. A description is at most 1,024 characters in the platform documentation and the open specification, and must say both what the skill does and when to use it; the help center article on uploading to claude.ai gives 200 characters, so keep the description short if the app is where it will live. Anthropic's authoring guide adds that the body should stay under 500 lines and that descriptions are written in the third person.

Here is the smallest valid skill, as Anthropic's open-source skills repository shows it:

---
name: my-skill-name
description: A clear description of what this skill does and when to use it
---

# My Skill Name

[Add your instructions here that Claude will follow when this skill is active]

Skills differ from prompts, Projects and MCP in one way each

Skills carry procedural knowledge that loads across conversations, where prompts last one conversation, Projects hold background knowledge that is always loaded inside the project, and MCP connects Claude to data. Anthropic's own comparison article from March 5, 2026 settles the pair that matters most: "MCP connects Claude to data; Skills teach Claude what to do with that data."

That sentence decides most of the arguments inside a paid team: the Monday report you keep re-explaining is a skill, the brand positioning you want remembered is a Project or a reference file, and reading the Google Ads account is an MCP server, which no skill replaces. Sorting that out early lets the team hand the Monday report to a skill, and Scalewell helps decide which jobs become skills and which data needs a connector.

Building blockCarriesLoadsGood for in a paid team
SkillA method, templates, scriptsWhen the request matches the descriptionWeekly readout, search terms review, pacing check
PromptOne-off instructionsEvery turn of one conversationA question you will ask once
ProjectBackground knowledgeAlways, inside that projectBrand voice, ICP notes, past decks
MCP serverA connection to a systemWhenever the connector is onReading Google Ads or Meta data, writing changes

Skills run in the Claude app, Claude Code and the API, but they do not sync

Custom skills work in three places as of November 2026, and a skill uploaded in one place is not available in the others; Anthropic's documentation says it directly: "Custom Skills do not sync across surfaces." In the Claude app you zip the skill folder and upload it under Customize, then Skills, on Pro, Max, Ultra, Team and Enterprise plans with code execution turned on, according to the platform documentation; Anthropic's help center lists the Free plan as well, so check the plan line in your own settings before you assume you need to upgrade. In Claude Code you drop the folder into ~/.claude/skills/ for yourself or .claude/skills/ inside a repository for the team, and the Claude Code documentation notes that a signed-in terminal also syncs skills from the claude.ai account since version 2.1.273, though not when it runs on an API key or a cloud provider. On the API you upload through the /v1/skills endpoint and the skill runs in a container with no network access.

Our decision rule for where to put a skill is short. Build it in Claude Code when it runs scripts against exported files or an MCP server you configured locally, and commit it to the repository so the team gets the same version. Upload it to the Claude app when the people using it are account managers who live in the chat window and the data arrives through a connector the admin set up. The API is for products and scheduled jobs.

Build skills in the order their output can be checked

The first skill a paid media team should build is the one whose output a human already checks every week, because a skill nobody verifies will drift and a skill everybody verifies will improve. We would rank the candidates on two questions: how often the job repeats, and how easy it is to say whether Claude got it right. Reporting scores high on both. A full account audit scores low on both, which is why it goes last even though it is the skill every vendor sells first.

This rule has a cost. The skills with the most obvious payoff, the ones that promise to find wasted spend across an account, are also the ones where Claude is most likely to produce a confident answer from thin data. Starting with reporting feels slow, and it builds the account context files that every later skill reuses. Anthropic's authoring guide points the same way: "run Claude on representative tasks without a Skill" first, document what it gets wrong, and write "just enough content to address the gaps."

A six-step staircase with a simple object on each step, rising toward the most complex task at the top.
The build order, ranked by how often the job repeats and how easy its output is to check: 1 weekly account readout, 2 search terms review, 3 budget pacing check, 4 launch checklist, 5 ad copy within platform limits, 6 account audit. The readout scores highest on both questions; the audit scores lowest on both, so it comes last.

Skill 1: the weekly account readout

The weekly readout skill turns a fixed set of numbers into a short written account of what changed, why it changed and what you will do about it, in the same order every week. It comes first because the team already argues about its format, which means the method is half written, and because the paid media lead reads every one, which means mistakes get caught.

weekly-readout/
├── SKILL.md
├── references/
│   ├── account-context.md     (conversion actions that count, brand terms, campaign naming, budgets)
│   └── readout-format.md      (the sections, the order, one real readout you were happy with)
└── scripts/
    └── compare_periods.py     (reads two CSV exports, returns deltas and flags beyond a threshold)

The SKILL.md body tells Claude which metrics to pull or accept, which period to compare against, which conversion actions count as pipeline and which are noise, and the section order of the readout. A real example sits in the format file, because Anthropic's own tutorial says "A real output you were happy with tells Claude more about your quality bar than any description of it." Arithmetic goes in the script, so Claude never computes a percentage change in its head.

A description that triggers well reads like this: "Writes the weekly paid media readout for a Google Ads or Meta account from exported data or an MCP connection. Use for the weekly report, the Monday readout, or what changed this week." That also fits under the claude.ai 200-character limit; in Claude Code or on the API you have room to add the other phrases your team says, such as a performance summary for a client or stakeholder.

Put together, the file reads like this. The shape is the one Anthropic's skill structure documents, frontmatter and then an Instructions section of numbered steps, with the reference files linked by relative path and the script named as something to run rather than read, which the authoring guide asks for. The week definition, the thresholds and the section names are our starting point and yours to change.

---
name: weekly-readout
description: Writes the weekly paid media readout for a Google Ads or Meta account from exported data or an MCP connection. Use for the weekly report, the Monday readout, or what changed this week.
---

# Weekly readout

## Instructions

1. Accept a campaign-level export, or pull one through the account's MCP connection, with spend, impressions, clicks, conversions by action and conversion value for the last full week (Monday to Sunday) and the week before it.
2. Count only the conversion actions listed as pipeline in [references/account-context.md](references/account-context.md). Report the others in one line as noise and never add the two together.
3. Run `python scripts/compare_periods.py this_week.csv last_week.csv` for every delta and quote its output. Do not compute a percentage change yourself.
4. Flag a change only when the script marks it beyond the threshold in account-context.md. Everything else goes in the table without comment.
5. Write the readout in the section order in [references/readout-format.md](references/readout-format.md): what changed, why it changed, what we will do, then the table of numbers.
6. Match the length and tone of the example readout in readout-format.md.
7. When a number you need is missing from the export, write "not in the export" in its place instead of estimating it.
8. End with the three decisions you want the paid media lead to make this week, each a complete sentence that names a campaign.

Skill 2: the search terms review

The search terms review skill reads a week of search terms, groups the ones that break the account's rules, and proposes negative keywords for a human to approve, without applying anything itself. It comes second because the rules are concrete and the output is a list you can check in ten minutes. The folder holds a SKILL.md with the review steps and output table, a references/negative-rules.md with the brand terms you protect, the competitor policy, the geographies and job titles that are out of scope, and the words that mean someone wants a free tool or a job, plus an export of existing negatives so the skill does not propose what is already there. The full method is in using AI to review search terms every week.

Skill 3: the budget pacing check

The pacing check skill compares month-to-date spend against the month's budget by campaign and names the campaigns that will overspend or underspend at the current rate. It comes third because the arithmetic is deterministic, so it belongs in a script, while the tolerances are a judgment call that should live in a reference file the team can edit. The folder holds SKILL.md, references/budgets.md with each campaign's monthly budget and tolerance, and scripts/pacing.py that takes spend to date, days elapsed and days remaining and returns projected spend and the gap. Anthropic's authoring guide calls this the low-freedom pattern: when "operations are fragile and error-prone" and "consistency is critical", give Claude a script to run rather than a method to follow.

Skill 4: the launch checklist

The launch checklist skill checks a planned campaign against the account's conventions before it goes live: the naming pattern, the UTM template, the conversion action it optimizes to, the geography, the audience exclusions and the landing page. It comes fourth because the mistakes it prevents are the expensive kind and the checks are binary, but it runs less often than the first three. The folder holds SKILL.md with the checklist in order, the naming convention and UTM template as reference files, and scripts/validate_names.py that tests proposed names and prints the failures. Ask for a pass or fail per check and nothing else, so a junior can run it before asking a senior to look.

Skill 5: ad copy drafting within platform limits

The ad copy skill drafts headlines and descriptions from a brief and the brand's voice, then checks every line against the platform's character limits and the claims the brand may make. It comes fifth because the output is creative and harder to judge, but the limits are fixed: Google's responsive search ads help page says headline fields "support up to 30 characters" and description fields "support up to 90 characters each", with 3 to 15 headlines and 2 to 4 descriptions per ad. The folder holds SKILL.md with the drafting steps and a review loop, references/brand-voice.md, references/approved-claims.md, and scripts/check_lengths.py that flags lines over the limit. Anthropic's guide recommends exactly that loop, run the validator, fix the errors and repeat, and says it "greatly improves output quality." For the prompts themselves, see a prompt library for paid media.

Skill 6: the account audit

The account audit skill comes last because it needs everything above it. A credible audit reads conversion definitions, naming conventions, budgets and negative rules, pulls several reports, compares periods and then judges what matters. Build it by composing the earlier skills; Anthropic says Claude "automatically identifies which skills are needed and coordinates their use" when several are installed. Keep the audit SKILL.md to the order of investigation and the output format, and point it at the reference files the other five already maintain.

Google ships official Google Ads skills, Meta and LinkedIn do not

Google publishes three official skills for the Google Ads API in the Agent Skills format, installed with one command, as of November 2026. The Google Ads API agent skills page lists google-ads-api-mcp-setup, which configures Google's Google Ads MCP server, google-ads-api-account-diagnostics, which diagnoses conversion loss, low lead volume and lost impression share through that server, and google-ads-api-quickstart for developers making a first API call. In Claude Code they install with npx skills add google/skills/skills/ads --agent=claude-code; the command installs Google's whole ads bundle, which also carries Data Manager API and Mobile Ads SDK skills you can ignore, and the repository is Apache 2.0 licensed with about 23,600 GitHub stars when we checked. For a model of what your own skill could look like, read the diagnostics skill: its SKILL.md names the MCP tools to call, separates workflows for conversion loss, impression share and low lead flow, and records details such as cost coming back in micros.

The server it pairs with is almost read-only: it can create assets and manage labels behind a confirmation step, and it cannot change bids, budgets, keywords or campaign status, so a skill built on it recommends changes without applying them. Set-up and authentication are covered in Google Ads MCP: how to connect Claude or ChatGPT to your account.

Meta publishes a hosted ads MCP server at https://mcp.facebook.com/ads with seven tool groups including reporting and ad creation, and opened it to any developer with a Meta app on July 16, 2026, but we found no official Meta skill in the Agent Skills format. LinkedIn's own MCP server is in beta and already in use as of November 2026, though not yet publicly documented, and we found no official LinkedIn skill. For both platforms your skill will run on data from a connector, an export or a third-party server.

Community packs are a different category. Corey Haines's marketingskills repository had about 59,100 stars when we checked, is MIT licensed and holds around fifty skills including ads and ad-creative. Its ads skill is a general method that tells Claude to label unavailable inputs as missing rather than estimating them, and what it cannot hold is your conversion definitions or your budgets. Treat a pack like that as a reading list for writing your own.

Writing the description decides whether the skill ever runs

The name and description are all Claude reads before deciding to open a skill, so a vague description means a skill that never triggers and a broad one means a skill that triggers on the wrong requests. Anthropic's own skill-creator skill, in the anthropics/skills repository, is candid about which way the failure runs: "currently Claude has a tendency to 'undertrigger' skills -- to not use them when they'd be useful. To combat this, please make the skill descriptions a little bit 'pushy'."

In practice that means listing the phrases your team says. A paid media lead rarely asks for "a performance summary"; they ask "what happened last week" or "why did CPL jump", so those words belong in the description. Anthropic's academy tutorial gives the test: open a new conversation, ask the way you normally would, and look for "Reading [your skill's name]" in Claude's thinking. When it is missing, the request did not match the name or description, and the fix is to change the description rather than prompt harder.

Claude Code adds one constraint: each entry's description is cut at 1,536 characters in the skill listing, so the documentation says to "put the key use case first."

Most marketing skills fail for the same three reasons

The skills that disappoint paid media teams fail in predictable ways, and none of the failures is about the model.

The skill has no data. A skill that says "audit the account for wasted spend" with no connector and no export will produce an audit anyway, because that is what it was asked to do. Imagine two hypothetical teams: Company A installs a fifty-skill pack, asks for a Google Ads audit in a fresh chat, and gets a well-structured document whose numbers Claude had to assume, while Company B writes a sixty-line readout skill, pastes in a CSV export, and gets a correct readout of one week. The second team is behind on features and ahead on trust, and trust is what earns a second week.

The skill is a prompt in a folder. When the whole SKILL.md says "you are an expert PPC strategist, analyze the following", it adds nothing Claude did not have. Anthropic's authoring guide says the default assumption is that "Claude is already very smart" and to add only context Claude lacks: your conversion actions, your naming convention, your thresholds, your format.

The skill was installed, not read. A skill can include scripts that run with whatever access Claude has, and Anthropic's documentation says to treat installing one like installing software, warning that "a malicious Skill can direct Claude to call tools or run code in ways that don't match the Skill's stated purpose." Read every file before you install a pack.

There is also a limit to how many hours skills will save, and the evidence is indirect. The State of PPC 2026 report, from PPCsurvey.com with 1,306 PPC professionals surveyed in late 2025, found that most individual contributors and freelancers save between one and five hours a week with AI, which the report calls "Useful, but not dramatic", and almost nobody reports saving 20 or more hours; 53% of practitioners with at least two years' experience said managing PPC campaigns is harder than two years ago, up from 49% two years earlier. DoubleVerify's 2026 survey of 2,020 marketing decision-makers put manual optimization at just under 10 hours a week, as reported by PPC Land. Neither figure measures skills, and we found no credible study that does. One to five hours a week is a real gain and a long way from the account running itself, which is the right frame for what six skills will do. For the wider picture, see what AI agents can realistically run in a paid media team.

Where to start this week

Pick the report your team argues about most, write down how a good version of it is built, and turn that into your first skill with the account context in a reference file beside it. Run it against a real export before you connect anything. If you would like help setting the method up around how your team already works, Scalewell does that with paid teams.

Frequently asked questions

Are Claude skills the same as Claude Code slash commands?

In Claude Code they are the same thing now. Anthropic's documentation says custom commands have been merged into skills: a file at .claude/commands/deploy.md and a skill at .claude/skills/deploy/SKILL.md both create /deploy. Skills add a folder for supporting files and frontmatter that controls whether you or Claude invokes them.

Can a Claude skill change bids or pause campaigns?

A skill changes nothing on its own; it can only act through a connector that allows writes. Google's official MCP server reads anything and, since November 2, 2026, can create assets and manage labels behind a confirmation step, but it cannot touch bids, budgets, keywords or campaign status, so a skill using it can recommend account changes without applying them. Meta's hosted server includes ad creation and management tools, so a skill paired with it can write, which is a reason to keep approval steps inside the skill.

Can Claude write the skill for me?

It can, and Anthropic recommends working that way. The authoring guide says "Claude models handle the Skill format and structure natively" and suggests finishing a task with Claude once, then asking it to capture the pattern as a skill. Review the draft for anything Claude already knows, test it in a fresh conversation, and bring what went wrong back to the next edit.

How long should a paid media skill be?

Shorter than most teams expect. A weekly readout skill that fits on two screens, plus an account context file and a format example, is a normal size, with detail in separate reference files Claude reads only when needed. Longer usually means the skill is explaining things Claude already knows, and Anthropic's 500-line ceiling is a limit rather than a target.

Do Claude skills work in ChatGPT, Codex or Cursor?

The format is portable, the installed skill is not. Anthropic published Agent Skills as an open standard in December 2025, and Google's Google Ads skills page gives install commands for Claude Code, Codex, Cursor and Antigravity from the same folder. Each agent still needs its own copy, and behavior depends on the model reading it, so test in every tool you use.