Claude Cowork Onboarding SkillBuilt the "Ripple Cowork Setup" skill that calibrates Claude to any Ripple role in about fifteen minutes, adopted company-wide and selected for Ripple's AI Demo Day.
Q1 2026
Context
When Cowork launched at Ripple, most people didn't know what it was, let alone how to get real value from it. The people who did figure it out treated it as a personal edge.
The obvious move when you're ahead on a new capability is to keep the advantage. I built the "Ripple Cowork Setup" skill on the opposite bet: the way to win when something like this arrives is to be the one who hands it to everyone, not the one who hoards it. It walks any teammate, not just technical or research roles, through a short guided intake and generates two files: a CLAUDE.md that auto-loads their full context every time they open their workspace, and a portable Global Instructions block that follows them across folders. Fifteen minutes once, and every session after starts calibrated instead of generic, the difference between output you have to double-check and output you can just use.
Outcome - The skill was adopted across the company and selected for Ripple's company-wide AI Demo, where the strongest proof came from the demo itself: prepping the presentation with Cowork, the assistant applied my own research methodology back at me, proposing persona-based testing before I asked for it.
Tasks & Actions
Built a guided intake flow that adapts its questions by role, so an engineer, a PMM, and a treasury ops analyst each get calibrated differently instead of filling out the same generic form.
Wrote fallback rules so the skill still produces a useful first draft with partial input, missing writing samples, or a first-time AI user, rather than blocking on incomplete answers.
Used the skill to prep my own presentation for Ripple's company-wide AI Demo under a three-day deadline. It proposed persona-based testing unprompted, running three synthetic profiles, an engineer, a PMM, and a treasury ops analyst, against its own output.
Caught and fixed real breakpoints those personas surfaced: inconsistent identity across responses, leaked instructions in copy-paste output, and drifting template structure.
Iterated the actual demo deck and video with the same tool, down to slide headlines and pacing.
