AI marketing stack: models, skills and tools sorted by job
Last verified
The AI marketing stack is a curated base of AI models, agent skills and MCP servers for paid acquisition and app growth. Each entry names the job it does, its licence, what it can change in your accounts and the date its facts were last checked against the vendor or repository.
114 entries: 36 models, 46 skill packs, 18 MCP servers and 14 tools and lists. Last verified .
Jobs and guides
- AI for ad copy: picking a model and testing the output
Pick an AI model for ad copy by tone, tool access and cost, then test it on your own brief. A media buyer's workflow, brief template and kill rule per variant.
- AI for static ad creatives
Image models for static Meta and TikTok ads compared on what matters in ads: text in the image, product accuracy, edit control, rights and vendor price.
- AI for video ads, UGC-style avatars and voice-over
Video models, avatar tools and voice engines for ad production, with host prices, commercial-use terms, AI disclosure notes and a cost-per-variant method.
- AI for market and customer research
Use AI to mine reviews, Reddit and competitor ads for hooks without invented facts: a research workflow with source links, verbatim quotes and a verify step.
- AI for marketing analytics and reporting
AI for marketing analytics depends on data access more than model choice. Connect GA4, ad platforms and payments, add sanity checks, keep numbers reproducible.
- AI agents in paid user acquisition: delegate, gate or keep
A delegation table for AI agents in paid user acquisition: which tasks to hand over, which need human approval, which to keep, plus logging and rollback rules.
- Meta Ads MCP: a safe setup for AI agents
Connect an AI agent to Meta ads through the official MCP server without risking spend: read-only first, scoped access, approval for writes, caps, log review.
- How to install and vet marketing skills
Install marketing skills with the skills CLI, then vet each one: read the licence, scan bundled scripts, check what it writes and pin the version you tested.
Skill base preview
| Name | Job | Skills | Licence | Tested | Checked |
|---|---|---|---|---|---|
| ua5er/user-acquisition-skills | Paid UA agents, Analytics, Landing pages | 15 | MITClass A | 2026-11-15 | |
| coreyhaines31/marketingskills | Ad copy, Landing pages, Analytics, SEO and GEO | 53 | MITClass A | 2026-11-15 | |
| AgriciDaniel/claude-ads | Paid UA agents, Analytics | 34 | MITClass A | 2026-11-15 | |
| autonnel/autonnel-skills | Landing pages, Analytics | 6 | Apache-2.0Class A | 2026-11-15 | |
| aaron-he-zhu/aaron-marketing-skills | SEO and GEO, Market research | 120 | Apache-2.0Class A | Not tested yet | |
| addyosmani/web-quality-skills | SEO and GEO | 6 | MITClass A | Not tested yet | |
| adkit/ads-skills | Paid UA agents | 11 | MITClass A | Not tested yet | |
| AgriciDaniel/claude-seo | SEO and GEO | 35 | MITClass A | Not tested yet |
Featured deep-dive
Jev for marketing: Claude + Jev on a real ad market. Jev in a Claude + Jev creative system, tested on 54 live ads: typed answers, a cost ledger, a pre-flight panel, install steps, limits and what it cannot prove.
The whole base is one JSON file, licensed CC BY 4.0: /ai/base.json. How entries are checked is on the method page.
What the base covers and how entries are chosen
The base holds 114 entries as of the check on 15 November 2026. Of these, 36 are AI models, and 22 of those have facts verified against the vendor. The other 78 are repositories: 49 in class A and 29 in class B. A further 11 class C repositories are withheld from the list. Three models are named without any facts (Seedream, Kling 4.0, Midjourney), so treat them as pointers and not as vetted entries.
An agent skill is a SKILL.md folder that gives a coding agent such as Claude Code, Codex or Cursor a procedure for one task. An MCP server is a program that lets the agent call outside tools, for example an ad platform API. Both can touch real accounts, so each entry records what it can write, what leaves your machine, and whether it ships scripts. The method page lists the selection rules and the class definitions.
Size is not a criterion. Packs of 100 to 1,000 skills exist, and most of that volume is near-duplicate topics. Depth shows up more often in packs of 6 to 50 skills, which is where the base spends its attention.
Jobs and guides
Entries are sorted by the job you need done, not by repository name. Typical jobs are launching a Meta test, auditing a landing page and measuring AI traffic. Each job page lists the models, skills and MCP servers for that job and what to test, so you can compare options against your own account setup, budget and risk tolerance.
Models are handled the same way. The useful question is which model suits a given brief at a given cost. A single winner across all jobs is not claimed anywhere in the base. A comparison should show the same brief run through each model, with the date of the run.
Read three columns before you install
Read three columns before you install anything: the licence, the write access and the date checked. A pick that can only read your ad account costs you little to try. A pick that can change budgets or pause campaigns needs approval gates and spend caps before it touches a live account.
Several MCP servers are read-only by design, including the TikTok ones found and at least one Klaviyo audit skill. The base treats read-only as a filter, and it is a sensible default for a first week with any new tool.
Skill base preview
The preview shows eight rows from the full skill list. Use it to judge the format before you open the list: job, source, licence class, write access and check date. Four packs have had a full read, a licence read at a pinned commit and a scan: ua5er/user-acquisition-skills, coreyhaines31/marketingskills, AgriciDaniel/claude-ads and autonnel/autonnel-skills. The notes on those four are drafts that the site owner has not reviewed yet, so read them as working notes.
Three packs in the base, ua5er/user-acquisition-skills, ua5er/web2web-bi-board and ua5er/dc-lite, are published by the site owner. They are marked as first-party so you can weigh them accordingly.
Four traits separate the stronger skills from the weaker ones:
- They state when not to use the skill, as autonnel does.
- They cap their claims by the inputs they received.
- They separate observation from recommendation, as claude-ads does.
- They enforce read before write for any action on an ad account, as mcp-ads and adloop do.
Weak skills are usually persona prompts such as "act as an expert media buyer", with no procedure, thresholds or inputs. They add little over a plain prompt. Many skills also depend on a project context file, so expect to write one about your product, audience and unit economics before the output becomes useful.
Using the base
The full base is one JSON file at /ai/base.json, licensed CC BY 4.0, so you can load it into your own tooling and filter it by licence class, write access or check date. Entries are re-checked on a monthly cycle, and the date on each row tells you how fresh its facts are.
If you want these tools wired into your own stack, Vlad Yasinski, a performance marketer and growth consultant in Warsaw, takes engagements covering BI and analytics, automated advertising operations, competitor intelligence on public data and platform terms, integration of workflows and AI agents, and hands-on paid acquisition.
See the featured model on a real job: Claude + Jev read 54 live ads, tagged them with typed answers and pre-flighted six example concepts for 1.5 cents.
See the real runSources
Each source shows the date it was last checked.
Show all 16 sources
- Agent Skills agentskills.io/home
- Anthropic claude.com/pricing
- Anthropic code.claude.com/docs/en/overview
- Cursor cursor.com/pricing
- GitHub github.com/ua5er/dc-lite
- GitHub github.com/ua5er/web2web-bi-board
- learn.chatgpt.com learn.chatgpt.com/docs/pricing
- GitHub AgriciDaniel/claude-ads: LICENSE
- GitHub autonnel/autonnel-skills: LICENSE
- GitHub coreyhaines31/marketingskills: LICENSE
- GitHub ua5er/user-acquisition-skills: LICENSE
- GitHub AgriciDaniel/claude-ads: README.md
- GitHub AgriciDaniel/claude-ads
- GitHub autonnel/autonnel-skills
- GitHub coreyhaines31/marketingskills
- GitHub ua5er/user-acquisition-skills
Want this applied to your own ad accounts? Start with a short audit.
See the service