Setup · 15 Minutes
Installing the Agent Pack
Nine skills, plain Markdown, no vendor lock-in. This is the full setup path my own team follows — from download to your first weekly revenue digest.
Download the packBook a call
Mentoring a young Ukrainian marketing graduate
What you need before you start
Fifteen minutes and three things:
A Claude subscriptionClaude Pro or Max (claude.ai) — or any agent framework that reads Markdown instructions. The skills are plain text; nothing is locked to one vendor.
Read access to your dataGA4, Google Ads, Microsoft Clarity, and your store platform (Magento, Shopify, or WooCommerce admin). Read-only is enough — the agents propose, humans change.
A named humanSomeone who reviews the approval queue. The pack is built so agents never touch a live store or ad account on their own.
The pack itselfDownload the zip — nine skills, each a folder containing one SKILL.md file, plus a README with the ground rules.
Installing the skills
- Unzip the pack. Inside skills/ you will find nine folders. Each folder is one agent skill — open any SKILL.md in a text editor and read it; there is no hidden code, just the operating procedure.
- Claude desktop or Cowork. Settings → Capabilities → Skills → Add skill → upload a skill folder (or the zip). Repeat for each skill you want — start with revenue-data-aggregator and ecommerce-growth-audit.
- Claude Code. Copy the folders into .claude/skills/ inside your project, or ~/.claude/skills/ to make them global.
- Connect your data. In Claude, enable the connectors you have (Google Analytics, Google Ads, and your browser for anything without an API). Exports work too — the skills accept CSV exports when live access is not set up yet.
- Run your first command. Ask: “Run the ecommerce growth audit on [store URL]” or “Build the weekly revenue digest for [client]”. The agent will ask for whatever access it is missing and mark anything unverifiable as NOT VERIFIED rather than guessing.
✋ The ground rules (read these first)
Baked into every skill in the pack:
- Platform orders outrank analytics; analytics outrank ad-platform claims — disagreements get reported, never averaged away
- Read-only until approved: agents propose, named humans approve each item
- Evidence or silence — every number traces to a reproducible report
- Change velocity caps: bids ±15% per week, budgets ≤20%, no pausing without sign-off, no under-powered tests
- Every run gets logged — time, cost, outcomes — so the system improves
A sensible first month
- Week 1 — establish truth. Run the revenue data aggregator on one store. Fix whatever the reconciliation exposes (it will expose something).
- Week 2 — audit. Run the growth audit and the ads account audit. You now have a scored baseline and a priced fix list.
- Week 3 — start the loop. Turn on the weekly optimization queue and the feed QA. Approve from the queue; measure the hours you got back.
- Week 4 — compound. Ship the first client report from reconciled numbers, plan the first tests with the CRO planner, and document what you learned.
Stuck on setup? My team at MageCloud installs and operates this system inside client engagements — book a call. For the philosophy behind the recursive skill system, read how Dennis Yu’s agents document and improve themselves.
© Paul Ryazanov · paulryazanov.com · MageCloud · Ecommerce CampBuilt in partnership with Dennis Yu · how this system was built
