Skip to content
AI marketing stack

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.

How the base works
How the base worksA marketing job is matched to models, skills and MCP servers; each entry passes a licence check, a safety check and a dated verification, and what passes becomes a small stack for that job.THE BASE, MATCHEDEVERY ENTRY PASSESLICENCESAFETYDATEVERIFIEDMarketing jobexample: ad copyModelstext, image, videoSkillsfor your agentMCP serversread-only firstYOUR STACKModelSkillMCP serverHow the base worksA marketing job is matched to models, skills and MCP servers; each entry passes a licence check, a safety check and a dated verification, and what passes becomes a small stack for that job.Marketing jobexample: ad copyModelsSkillsMCP serversLICENCESAFETYDATE VERIFIEDYOUR STACK FOR THIS JOBModelSkillMCP server
Illustrative flow, simplified.
Start from the job. The base matches models, skills and MCP servers to the job, and only entries that pass the licence, safety and date checks reach your stack.

114 entries: 36 models, 46 skill packs, 18 MCP servers and 14 tools and lists. Last verified .

Jobs and guides

Skill base preview

Eight agent skill packs from the base, licence class A
NameJobSkillsLicenceTestedChecked
ua5er/user-acquisition-skillsPaid UA agents, Analytics, Landing pages15MITClass A2026-11-15
coreyhaines31/marketingskillsAd copy, Landing pages, Analytics, SEO and GEO53MITClass A2026-11-15
AgriciDaniel/claude-adsPaid UA agents, Analytics34MITClass A2026-11-15
autonnel/autonnel-skillsLanding pages, Analytics6Apache-2.0Class A2026-11-15
aaron-he-zhu/aaron-marketing-skillsSEO and GEO, Market research120Apache-2.0Class ANot tested yet
addyosmani/web-quality-skillsSEO and GEO6MITClass ANot tested yet
adkit/ads-skillsPaid UA agents11MITClass ANot tested yet
AgriciDaniel/claude-seoSEO and GEO35MITClass ANot tested yet

See all skill packs

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 run

Sources

Each source shows the date it was last checked.

Show all 16 sources
  1. Agent Skills agentskills.io/home
  2. Anthropic claude.com/pricing
  3. Anthropic code.claude.com/docs/en/overview
  4. Cursor cursor.com/pricing
  5. GitHub github.com/ua5er/dc-lite
  6. GitHub github.com/ua5er/web2web-bi-board
  7. learn.chatgpt.com learn.chatgpt.com/docs/pricing
  8. GitHub AgriciDaniel/claude-ads: LICENSE
  9. GitHub autonnel/autonnel-skills: LICENSE
  10. GitHub coreyhaines31/marketingskills: LICENSE
  11. GitHub ua5er/user-acquisition-skills: LICENSE
  12. GitHub AgriciDaniel/claude-ads: README.md
  13. GitHub AgriciDaniel/claude-ads
  14. GitHub autonnel/autonnel-skills
  15. GitHub coreyhaines31/marketingskills
  16. GitHub ua5er/user-acquisition-skills

Want this applied to your own ad accounts? Start with a short audit.

See the service