Agent skills

A Claude Code skill that teaches your agent the whole loop: set up, write variants, read results, promote.

.claude/

skills/

conversioncat/

SKILL.md when to use it, the workflow, the rules

references/

writing-variants.md one idea per variant, what to test first

reading-results.md p(best), uplift intervals, when to promote

onchain-goals.md wallet linking, transfer, call, hold

What ccat skills install copies into your project. Claude Code loads it when the task matches.

Install

ccat init installs it. To add or refresh it alone:

ccat skills install

What it knows

When to use ConversionCat: improving a conversion rate, A/B or multivariate tests, attribution, onchain goals.
The workflow, step by step, with the exact commands and the --json output it should read.
How to write a variant: one idea per variant, headline first, then the call to action, then proof.
How to read results: no call under 300 visitors per variant, promote at p_beat_control 0.95 with an uplift interval above zero.
The rules: never print or commit CCAT_KEY, keep the control untouched while a test runs.

Other agents

The skill is plain Markdown. Codex, Cursor or any agent that reads files can use SKILL.md as instructions; every command it calls prints JSON with --json.