claude-ads (AgriciDaniel): paid ads audit skills reviewed
Last verified
claude-ads is an MIT-licensed skill pack by AgriciDaniel that audits paid ad accounts on Meta, Google, TikTok and other platforms through a conductor skill and per-platform audit agents. It keeps observations apart from recommendations. Install it with npx skills add AgriciDaniel/claude-ads and review its bundled scripts first.
Facts
- Repository
- AgriciDaniel/claude-ads
- Publisher
- Third party
- Licence
- MIT (class A). File: LICENSE. Read on 15 November 2026. Commit fc3908e786c318776df89a3a098333da377364be.
- Skills
- 34
- Last push
- 1 November 2026
- Archived
- No
- Read depth
- Full read
- Last verified
- 15 November 2026
- Install
/plugin marketplace add agricidaniel/claude-ads /plugin install claude-ads@ai-marketing-hub-claude-ads
Safety
- Writes to
- Ad account
- Approval step before a write
- Yes
- Data leaves the machine
- Landing page fetches to user-supplied URLs through a guarded fetcher; image briefs go to Gemini, OpenAI, Stability or Replicate only when you configure a provider key; a maintainer script reads the GitHub API.
- Scripts in the repository
- 131
- scan-skill.py run
- 15 November 2026, commit fc3908e786c318776df89a3a098333da377364be
- Notes
- Account audits for agencies and in-house teams across Google, Meta, YouTube, LinkedIn, TikTok, Microsoft, Apple, Amazon, Reddit, Pinterest, Snapchat and X. A conductor routes each request to one skill, platform workers return structured findings, and a deterministic scorer turns them into health and evidence coverage.
Skills worth starting with
- ads (conductor): Operating contract: evidence order, mutation gate, untrusted-data rule and scoring boundaries for every command.
- ads-audit: Multi-platform audit that writes a JSON run bundle first and labels the result complete, provisional, partial or insufficient.
- ads-meta: Meta controls for Pixel and CAPI, attribution, creative fatigue, structure and policy, with a cold-start contract for new accounts.
- ads-attribution: Refuses to add incompatible conversion windows and reports sources side by side until definitions match.
- ads-test: Experiment design with randomisation unit, guardrails, pre-registered stop rules and no peeking.
- ads-math: CPA, ROAS, MER, break-even and LTV:CAC with formulas, inputs and downside and upside cases.
Operator note
The most careful ads audit pack in this set. It treats every export, page and ad text as untrusted data, keeps account writes behind a mutation gate that wants an exact before and after plan plus owner approval, and refuses permanent deletion. Each control ends in pass, fail, unknown or not applicable, and unknown is never guessed. The cost of that care: on an ordinary export most controls come back unknown, because the catalogue holds no thresholds, so you need Events Manager data, daily rows and your own targets before it says much. Scoring and reports run on a Python core the installer sets up, so install it as a Claude Code plugin rather than copying skill files. Use it for agency-style audits across several platforms where evidence and an audit trail matter more than a fast opinion.
Test note
Draft, not yet reviewed by the site owner. Run on 2026-11-15 in Claude Code with the ads-meta skill.
Input: Sample data, invented for this test: a 30-day Meta ad set export for an ecommerce store, EUR, Europe/Warsaw, Purchase objective, four prospecting ad sets with 18, 11, 8 and 4 purchases, three of them in 'Learning limited'. Pixel events present; no Events Manager data, no Event Match Quality, no CAPI evidence, no daily rows, no owner thresholds. Account and Pixel active since 2025.
Result: Following ads-meta by hand (the Python scorer was not run under the no-execution rule): the cold-start check classed account and Pixel as not cold and conversion maturity as unknown without lag data. Seven controls produced Finding v1 objects: M01 Pixel installed pass (high); M13 learning phase status fail on Meta's own 'Learning limited' field (medium); M02 CAPI, M03 deduplication, M04 Event Match Quality, M11 fragmentation and M28 creative fatigue unknown, each with the missing input named and a recovery hint. No health score, because the skill forbids computing scores in the prompt and evidence coverage was under 60 percent.
What worked: It refused to invent a fragmentation threshold or a fatigue rule, kept observation and diagnosis apart, and listed exactly which inputs would turn each unknown into a verdict. The cold-start contract stopped a new-Pixel assumption.
What broke or needed fixing: Five of seven controls came back unknown on a normal export, so a buyer gets a data request list more than a diagnosis. The 50-events learning rule that most buyers expect is absent from the catalogue by design. The value of the pack depends on running its Python runtime, which a skill-only install does not give you.
Use cases
- Paid UA agents: Run /ads audit meta on an ad set export plus Events Manager screenshots; get Finding v1 results per control with evidence, confidence and missing inputs.
- Analytics: Give Meta 7-day and Google 30-day conversion reports; ads-attribution refuses to sum them and returns a side-by-side table until windows and definitions match.
What to check before you install
- Gather your inputs first. The control catalogue carries no thresholds, so on a typical export most controls return unknown. Prepare Events Manager data, daily rows and your own targets, or expect thin output.
- Confirm the host. Only Claude Code is verified. README marks Codex, Gemini, Cursor, Windsurf and Goose as experimental.
- Install it as a plugin and check that install.sh builds the managed venv for claude_ads_core. As plain skill files, the agent returns findings JSON but no score or report.
- Keep write access off. Ads-launch and ads-optimize change an ad account only with --apply and owner approval of the exact before and after plan. Read that plan before approving.
- Budget for context: the conductor, thinking framework and control catalogue load before the first finding. Image briefs leave only if you set a provider key.
Alternatives
- ua5er/user-acquisition-skills has 15 skills for app and subscription paid UA, covering kill and scale rules, Meta launch and learning audits, paywalls and unit economics. It is published by the site owner.
- coreyhaines31/marketingskills has 53 skills across CRO, copywriting, paid ads, tracking, SEO and lifecycle email, sharing one product marketing context file.
- autonnel/autonnel-skills has 6 funnel skills for work after the ad click, including landing page audits, upsell chains and server-side conversion tracking. One skill needs an Autonnel instance.
Sources
Each source shows the date it was last checked.
- GitHub AgriciDaniel/claude-ads
- GitHub AgriciDaniel/claude-ads: LICENSE
- GitHub AgriciDaniel/claude-ads: README.md
Want your own AI and acquisition setup checked for gaps and risks? Start with a growth audit.
See the growth audit