<aside> 🟢
+10% Revenue Per Visitor in 90 days. Or we refund the engagement.
</aside>
One pass through the whole loop: research, hypothesis, design, build, QA, launch, read the result, roll the learning forward. Each stage gives you the exact Claude move and the connector doing the work. Copy the prompts. Steal the cadence.
This is the cornerstone playbook. The others go deep on one stage. This one strings all of them into the program in order, so you can sit down on a Monday and start.
The whole thing runs on one pattern.
Claude (the brain) + a connector (the hands) = an outcome that used to need a team.
A connector is either an MCP server (think "USB-C for AI", one standard plug that lets Claude read your data and take actions inside a tool) or a CLI (a command-line tool Claude drives directly). You don't write code. You just know which plug goes where and what to type.
Most people use AI like a vending machine: open a tab, paste a prompt, copy the output, repeat. Your CRO program can't run that way. A program is a loop, and the value is in the loop closing. The gap that matters is the one between "I asked AI for some test ideas" and "Claude pulled my GA4 funnel, found the mobile leak, scored the backlog, wrote the hypothesis, launched the variant in Convert, read the result at significance, and appended the learning to the test log so the next test is smarter."
That second sentence is the program. Here is how you build it.
| Stage | What happens | Claude move | Connector / tool |
|---|---|---|---|
| 1. Research | Assemble the evidence stack | Reads screenshots, funnel, heatmaps, reviews | GA4, Clarity, Apify, Shopify |
| 2. Audit + prioritize | Score the page, rank the backlog | Scores 1-5 per dimension, then ICE | (same inputs) |
| 3. Hypothesis | Turn a finding into a falsifiable bet | Writes strict-format hypothesis | none (reasoning) |
| 4. Design | Spec the variant + size the test | Variant brief + sample-size math | none (reasoning) |
| 5. Build | Get a dev-ready variant | Figma to code, or code direct | Figma MCP, Claude Code |
| 6. QA + launch | Set up the experiment, gate on human check | Builds it in the testing platform, returns preview URLs | Convert MCP |
| 7. Analyze | Read significance, segment, verdict | Sig check + segment + ship/kill/iterate | Convert MCP |
| 8. Roll forward | Document the learning, feed the next test | Writes the post-test analysis, logs it | Notion / Airtable MCP, Slack |
The arrow from stage 8 back to stage 1 is the part almost everyone skips. It is also the part that compounds.
Before the loop runs, three things have to exist. Skip these and you are back to the vending machine.
Open Claude Desktop if you are new. Connectors install from a marketplace in a couple of clicks, no config files, and it is best for chat-driven work: research, survey analysis, reporting, the audit.
Move to Claude Code for local execution and the build/launch recipes. It is stronger on long reasoning, file handling, and code. Graduate here the week you want to run the variant build or the launch flow. Desktop and Code share configuration and skills, so moving up is clean.