claude-ads Runs Evidence-Backed Paid-Media Audits Across 12 Ad Platforms

Paid-media reviews usually mean screenshots, spreadsheets, and gut feel. claude-ads is a 9k-star, MIT-licensed skill that gives Claude Code a structured operations layer for twelve ad platforms: authorized exports or account reads go in, and evidence-backed audits, plans, and versioned reports come out.

Why This Skill Matters

The coverage is broad but shallow coverage is not the point — each platform gets a focused skill, an audit worker, a control reference, a capability declaration, and a testable routing surface, and the README names the segment split: Google, Meta, YouTube, LinkedIn, TikTok, Microsoft Advertising, Reddit, Snapchat, and X Ads on the search/video/social side; Apple, Amazon, and Pinterest Ads for commerce and retail media.

The safety model is the differentiator. Everything is read-only by default, and a live change stays disabled until it passes approval, idempotency, verification, audit, and rollback gates for the exact platform and operation. Permanent deletion is not supported in v2. That makes it usable on real client accounts, not just demo data.

Installation

Claude Code is the canonical runtime. Use the native plugin flow from the README:

/plugin marketplace add agricidaniel/claude-ads
/plugin install claude-ads@ai-marketing-hub-claude-ads

In a plugin install the commands are namespaced — /claude-ads:ads — while standalone installs use plain /ads. If you added the marketplace before v2.0.0, remove the stale local alias agricidaniel-claude-ads first, then re-add and install; the README documents this exact sequence.

Prefer a local clone? The standalone path is:

git clone https://github.com/AgriciDaniel/claude-ads.git
cd claude-ads
bash install.sh --source=local

Managed dependencies target CPython 3.11 and 3.12; use --no-deps for a skill-only install. Browser capture needs an operator-installed Playwright browser, and PDF rendering needs WeasyPrint and Pango system libraries — both listed in the repo's external runtime dependency manifest.

Real Workflow: Audit a Google Ads Account

1. Describe the client once — KPIs, privacy constraints, guardrails:

/ads setup

2. Run a full audit, or scope it to one platform with the shortcut:

/ads google

3. Read the results. Every control scores pass, fail, unknown, or not_applicable. Coverage of 80% or more is graded, 60–79% is provisional, and below 60% is insufficient evidence — the README's scoring rules, so you know when an account's data is too thin to trust the health score.
4. Ask for improvements without touching the account:

/ads optimize --draft

The --draft flag produces an evidence-backed change plan; nothing is applied. /ads report renders the validated JSON run bundle into Markdown, HTML, or optional PDF — renderings of the same canonical data, not separately written documents.

Tips

• The capability manifest in control-plane/manifests/ is the authoritative record for what can read and write live — check it before assuming an operation is available.
• A failed platform makes the run partial and excludes it from portfolio scoring; the README is explicit that partial runs are never silently presented as complete audits.
• Platform shortcuts exist beyond Google: /ads meta, /ads amazon, and /ads reddit route to the matching platform audit.
• /ads experiment designs and reads out controlled tests; /ads monitor covers pacing, delivery, tracking, fatigue, policy, and performance.
• The repo ships from two homes: this public MIT release, and a community mirror for AI Marketing Hub Pro members with early access — the README's note distinguishes them.

When Not to Use This

This is an operations and audit layer, not an autopilot: if you want an agent to autonomously move budgets, the gate model will slow you down by design. It also works from authorized exports or account reads — without access to account data, you get planning workflows, not evidence-backed audits.

See the leaderboard for more skills.