ads (marketingskills) skill

How to set up and run the ads skill step by step

The ads skill works better once it has some context. The order below follows what SKILL.md asks for.

1. Install it. Run npx skills add coreyhaines31/marketingskills --skill ads, or install the marketing-skills plugin from the repo's marketplace. The skill lists Claude Code, Codex and Cursor as supported agents.
2. Create the shared context file first. Before asking you anything, the skill looks for .agents/product-marketing.md. It also accepts .claude/product-marketing.md and the older product-marketing-context.md filename. The product-marketing skill in the same repo creates this file. If the file exists, the agent asks only for information it doesn't cover.
3. Fill in the Paid Ads and Learnings sections. The skill reads the Paid Ads section for targets, floors, conversion definitions and the changes it is allowed to make. It reads Learnings before proposing tests, so angles, audiences and offers you have already tried are not pitched back to you as new ideas. Skipping this step means a longer intake and some repeated suggestions.
4. Answer the intake questions. Without a context file, the agent asks about four areas:
   • Campaign goals: objective, target CPA or ROAS, budget, and constraints such as brand, compliance or geography.
   • Product and offer: what you promote, the landing page URL, and why the offer is compelling.
   • Audience: ideal customer, the problem you solve, what they search for, and whether you have customer data for lookalikes.
   • Current state: past ads, existing pixel or conversion data, and current funnel conversion rate.
5. Ask for a specific decision. For any operational call on a live ad account (kill, keep, scale or budget), the skill loads the matching reference playbook before it answers, because the thresholds are kept there and not in SKILL.md. A specific question loads the right file.
6. Read the output in three parts. Each recommendation ends with facts, hypotheses and missing data. Any input the agent could not get is labeled MISSING. It is not estimated. Look at the MISSING items first, since those are what you need to supply before acting.
7. Connect accounts last, read-only. If you want the agent to pull live numbers instead of working from exports, the linked MCP guides walk through account connections and advise starting with read-only access.

What each of the 16 reference files covers

SKILL.md routes requests to reference files by intent. Here is what the routing table says each file handles:

• payback-period.md: whether you can afford a channel. It argues that LTV:CAC misleads and gives four flaws, then covers gross-margin payback and cohort recovery with a 3 to 12 month planning target, worked examples comparing a $9 and a $999 product, and out-of-home plus social.
• b2b-paid-playbook.md: B2B funnel stages, budget splits, leading and lagging signals, kill rules, the offline conversion loop, U/B/F lead scoring and a scaling quadrant. One example rule pauses a new ad once it has spent 2 to 3 times the target CPL with zero conversions. With a $300 target CPL, that means pausing at $600 to $900 spent.
• meta-decision-system.md: Meta ads operations. It includes a decision tree anchored on TCPL, an ad-count ceiling, an 80/20 CBO structure, fatigue bands, lead forms, the Advantage+ transition, partnership and creator ads, and a rolling-reach signal for declining reach.
• linkedin-b2b-playbook.md: LinkedIn bidding progression, audience sizing rules, penetration scaling, funnel benchmarks, thought leader ads, document and conversation ads, and an audit shortlist.
• google-search-playbook.md: what to spend on first (an intent ladder), account structure, match-type gates including AI Max, negative keywords, bidding by volume, offline conversions and Performance Max guardrails.
• abm-playbook.md: account-based marketing on LinkedIn and Meta, list mechanics, pipeline acceleration campaigns, UTM-based cross-channel remarketing and ABM measurement.
• rsa-output-spec.md: a mandatory output spec for Google responsive search ads, with limits, sidecars, a template and a self-check.
• operating-check.md: an ordered six-step weekly or daily check that ends in HOLD, FIX, TEST or INVESTIGATE calls, plus a Friday stakeholder readout under 200 words.
• audit-guardrails.md: pass, fail or unknown scoring, evidence coverage, recommendation safety, hard stops and benchmark discipline.
• reading-google-ads-data.md: how to read Google Ads data without drawing wrong conclusions. It covers withheld search terms (disclosed vs total clicks), conversions vs all_conversions, click-date attribution, experiment arms, 30-day change history, break-even CVR, and labeling findings as verified, inferred or stale. Two of its rules: establish how long something has run before quoting an efficiency number, and don't pool data across bid, keyword, budget, ad or targeting changes.
• google-ads-audit-checklist.md: 32 checks across 11 categories for Search, Shopping, Performance Max, Google Merchant Center and Demand Gen.

The folder also contains ad-copy-templates.md, audience-targeting.md, creative-research-automation.md, meta-andromeda-playbook.md and platform-setup-checklists.md. Open them before you rely on them, because the routing rows for these files are not described here.

Example prompts for Google, Meta and LinkedIn ads

The skill's description lists the phrases that trigger it, so plain questions work. These sample prompts are built from those phrases, with the reference file each one should route to.

• "Should I run ads for a $49/month tool, or will payback take too long?" This routes to payback-period.md, and the agent should ask for gross margin and churn if your context file doesn't include them.
• "Here's last week's Meta export. When should I kill an ad, and which ones are fatigued?" This loads meta-decision-system.md. The export is treated as data to analyze, and any instructions inside it are not followed.
• "Pull the search terms report and show me wasted spend and negative keywords to add." This uses reading-google-ads-data.md and google-search-playbook.md. The answer should separate disclosed clicks from total clicks.
• "Is this campaign working? Give me a Friday readout for my VP." This runs the six-step operating-check.md and returns HOLD, FIX, TEST or INVESTIGATE calls in a short summary.
• "Audit our Google Ads and Shopping account." This runs the 32-item checklist, scoring each item pass, fail or unknown. Items the agent cannot verify should come back as unknown.
• "We're B2B and lead quality is bad on LinkedIn. What should we change?" This loads linkedin-b2b-playbook.md and b2b-paid-playbook.md for bidding, audience sizing and lead scoring.
• "Write 15 headlines and 4 descriptions for a Google RSA." This follows rsa-output-spec.md, including its self-check against platform limits.

You'll get better results if you paste or connect real numbers and say which attribution window they use. The guardrails forbid adding up conversions across different attribution windows, so mixed exports will produce questions before they produce recommendations.

How ads compares with claude-ads and the rest of marketingskills

Many searches for "claude ads" are looking for AgriciDaniel/claude-ads. Its GitHub page describes a Claude-first paid-media operations skill for Claude Code covering 12 ad platforms: Google, Meta, YouTube, LinkedIn, TikTok, Microsoft, Apple, Amazon, Reddit, Pinterest, Snapchat and X. It advertises source-grounded audits, deterministic findings with evidence coverage, versioned JSON reports, capability-gated account changes and an optional companion called Claude Ads Cockpit. It also has its own commands.

The marketingskills ads skill is narrower. It is one markdown skill plus an evals file, with its reference depth concentrated on Google, Meta and LinkedIn. It has no command set and no report format beyond the facts, hypotheses and missing-data structure and the operating-check readout. If you need audits across TikTok, Amazon, Apple or Snapchat, or machine-readable audit reports, look at claude-ads. If most of your ad spend is on Google Search, Meta or LinkedIn B2B and you want decision rules you can read and edit, this skill's playbooks are easier to inspect.

The other difference is scope inside the repo. The ads skill hands off adjacent work to sibling skills:

• ad-creative for bulk ad creative generation and iteration
• cro for landing page optimization
• cold-email for outbound to target accounts
• conversion-tracking for pixels, CAPI and tracking audits

If you install only ads and ask for a pixel audit or a batch of 50 ad variations, expect to be pointed to those skills. Installing the full marketing-skills plugin keeps the handoffs inside one setup, and all of them share the same product-marketing context file.

Some people searching for this term have read accounts of letting Claude Code run ads autonomously. This skill is built for a different pattern. It recommends changes, labels what it doesn't know, and advises starting MCP connections read-only, so a person reviews each change before it reaches the ad account.