Twelve named, reusable Claude capabilities that run a CRO program end to end. Each one is a recipe I actually run: what it does, the tools it touches, the outcome, and a starter prompt you can paste today.

I think about Claude the way I think about a hire. A generalist who pastes one prompt at a time is a temp. An operator who has built a set of repeatable plays, wired into the store, the analytics, the ad accounts, and the design tools, is someone you can hand a whole function to.

A "skill" here is the second thing. It has a fixed shape: an input, a stack it touches, an outcome, and a prompt you keep and improve. Claude is the brain. The connectors and CLIs are the hands. The skill is the play you run over and over until it is muscle memory.

Everything below comes from the same program we run for $50M to $300M DTC Shopify brands. 1,000+ A/B tests across 150+ brands, 4 to 8 tests per brand per month. None of these skills is theoretical. Each one maps to a step we already do by hand, only faster.

One note before you start. A skill is only as good as the context behind it. Two things make every prompt below work harder: a reusable prompt library, and a brand voice file written once. Wire Claude into your data after those two exist, not before. Most of the value here is reusable context plus disciplined human review, not autonomy.

How to read this page

The 12 skills follow the actual CRO loop, in order:

1. Research and evidence (skills 1 to 3): figure out what is broken and why.
2. Hypothesis and design (skills 4 to 7): turn evidence into a ranked, designed, dev-ready test.
3. Launch and QA (skill 8): ship it without writing code yourself.
4. Read and learn (skills 9 to 10): decide if it won, segment the lift, write the learning.
5. Compounding layer (skills 11 to 12): make the whole thing run itself and stack across tests.

The first three skills feed the next four. The launch skill ships what design produced. The analysis skills close the loop and feed the next round of research. Run them in order on your first cycle. After that they overlap, because a real program always has tests sitting in research, in design, and in analysis at the same time.

Each skill below has the same four parts: what it does, the stack it touches, the outcome you walk away with, and a starter prompt. Paste the prompt, swap the brackets, go.

The connectors you wire in once

Most of these skills lean on a connector. A connector is either an MCP server (think USB-C for AI) or a CLI. You set them up once and every skill that needs them just works. The ones that matter for CRO:

A few setup commands you will reuse: