ai-marketing-claude (AI Marketing Suite for Claude Code)
A marketing skill pack for claude-code: 15 skills, 5 subagents, 4 Python scripts and 6 document
templates, MIT, installed with a piped shell script and driven from slash commands like
/market audit <url>. 2,686★ / 712 forks / 6 open issues / 37 watchers (2026-09-27 snapshot,
was 2,451★ / 692 forks / 7 open issues / 35 watchers on 2026-08-17). The third domain instance
of agent-skills in the marketing vertical, after agentic-seo-skill and claude-skills-ppc.
T1 in the ladder's sense — it is the primary artifact, the project speaking for itself. That is a statement about provenance, not quality, and the two come apart unusually far here.
What is actually in it
The repository is 37 files and 132 KB, and the tree was read rather than the README believed:
skills/— 14 directories, each oneSKILL.md: audit, ads, brand, competitors, copy, emails, funnel, landing, launch, proposal, report, report-pdf, seo, social. The advertised 15th ismarket/SKILL.md, the top-level dispatcher, so the headline count includes the router.agents/— 5 subagent definitions (content, conversion, competitive, technical, strategy), described as running in parallel.scripts/— 4 Python files:analyze_page.py,competitor_scanner.py,social_calendar.py,generate_pdf_report.py.reportlabis an optional dependency for the PDF.templates/— 6 markdown documents (three email sequences, a content calendar, a launch checklist, a client proposal).install.sh/uninstall.sh,requirements.txt, a banner SVG.
So the pack is almost entirely prompt text. Everything it claims to do — audit a website, score it across six dimensions, produce competitive intelligence, generate a client-ready PDF — is instructions to the model, with four small scripts beside them.
The comparison that makes it legible
agentic-seo-skill is the same form factor in the neighbouring vertical: 16 sub-skills, 10 agents,
and 89 scripts the project calls evidence collectors — pagespeed.py, crawl_audit.py,
validate_schema.py — each fetching a fact the model then reasons over. This pack has four.
That is the whole difference and it is a design choice, not an oversight. One pack sends the agent to measure the page; the other sends the agent to look at it and write the report. Both emit a document that reads like an audit. verification-loops-skills and the verification thread in synthesis hold why that gap matters: a skill without a checkable observation produces output whose confidence is uncorrelated with its accuracy, and the marketing deliverable — a scored, client-facing PDF — is exactly the artifact where an unverified score is most likely to be believed.
Nine minutes, then nothing
The metadata is the sharpest thing here. The repository was created 2026-03-02 01:09 UTC and last
pushed 2026-03-02 01:18 — nine minutes apart — and has not been touched in the nearly seven
months since (confirmed again on 2026-09-27: still one commit, e5aa0ea4, same push timestamp),
while accumulating 2,686 stars, 712 forks and 6 unanswered issues.
Nine minutes between creation and final push means the content was assembled elsewhere and landed in one go, which is ordinary. The silence afterwards is not. A 27% fork-to-star ratio (712/2,686, was 28% on 2026-08-17) is far above the norm for a tool repository and is what a pack looks like when people take a copy rather than depend on it — the ratio held steady while both numbers grew, so it is not an early-life artifact that would settle with more attention.
This is a clean specimen for the thread the spoke has been building. agentic-awesome-skills already recorded that a star count measures attention, and there it sat against 2,464 commits of visible maintenance. Here the same signal sits against one push. Whatever 2,451 stars measure, it is not that anyone found the pack correct — the pack has never been corrected.
The install path
curl -fsSL https://raw.githubusercontent.com/.../install.sh | bash is the documented method. For a
corpus that holds agent-skills-can-be-harmful — 307 hand-labelled failures caused by skills
themselves — piping an unread script from an unmaintained repository into a shell, to install
instructions an agent will then execute with the user's tools, stacks two trust decisions the user is
unlikely to be separating. Nothing here suggests bad intent; the point is that the delivery mechanism
gives the user no moment to look. uninstall.sh is at least shipped alongside.
What it is good evidence of
Not that the skills work. It is good evidence of the form factor's spread into non-engineering verticals and of what a marketing-shaped pack contains when someone builds one: a router skill, a fan-out of specialists, a small pile of scripts and a set of client-facing document templates. The templates are the tell about the audience — proposals and launch checklists are agency deliverables, and the README names agency builders and solopreneurs. This is agent-skills packaged as a service business in a box.
The same author's other pack answers most of this (2026-08-31)
geo-seo-claude, by zubair-trabzada, was routed here on 2026-08-31: MIT, 9,877★ / 1,531 forks, created 2026-02-18 and still being pushed the day it was ingested. It is the same form factor in the GEO/AI-search vertical, and set beside this pack it changes what the findings above are findings about.
| This pack (2026-03-02) | geo-seo-claude (2026-08-31) |
|---|---|
| Created and last pushed nine minutes apart, untouched since | Six months of pushes, still active |
| 4 Python scripts, none collecting evidence | 6 scripts, four of them evidence collectors (page fetch, citability scoring from measurable text properties, robots.txt crawler checks, llms.txt validation) |
| Six-dimension score, no published rubric | 16 KB methodology doc with per-point criteria — and an explicit "these weights are not derived from a controlled study" |
| No output sample | examples/ ships a real client audit, JSON, proposal and PDF |
| 28% fork-to-star ratio | 15% |
Two consequences for this page. First, "nine minutes, then nothing" is a fact about this repository, not about its author — he maintains the other one. The abandonment signal belongs to the artifact, which is the more useful reading anyway. Second, the "scored client PDF rests on no checkable observation" verdict below still stands here and does not carry across: the newer pack hangs its score on fetched facts. The design choice this page called "not an oversight" was revised by the same person within six months, in the direction of measuring.
Recorded weaknesses
Everything descriptive above is either the project describing itself or the file tree, and nothing has been run: no output sample, no evaluation, no third-party review, and no verifier of the kind skillsbench defines. The six-dimension scoring the README advertises has no published rubric in the repository root and no measurement behind it. Star and fork counts are a 2026-09-27 snapshot (refreshed from 2026-08-17; commit history unchanged).
A fifth instance adds a genuine zero (2026-10-07)
claude-marketing-research-skill extends the evidence-collector spectrum this page set up —
0 scripts in claude-skills-ppc (an article), 4 here, 89 in agentic-seo-skill — with a true
zero: one SKILL.md, no scripts directory, competitor data sourced from "web search" rather than
an author-written fetcher. It's also small enough (55★, 3 commits) that its install-command typo
(wrong repo name) and unfilled "[Your Name]" credit are still visible — a cleaner look at a pack
before the single-push-then-silence pattern this page describes, rather than after it.
Related
agent-skills · geo-seo-claude · zubair-trabzada · agentic-seo-skill · claude-skills-ppc · agentic-awesome-skills · claude-marketing-research-skill · agent-skills-can-be-harmful · skillsbench · verification-loops-skills · claude-code · synthesis