AI in the team, checked by people
I moved OMNE's analysts and designers onto Claude, built Claude skills for their day-to-day work, and keep clear rules on when to trust what it produces. The point is fewer gaps between what got decided, designed and built.
- Team
- 15 analysts and designers
- Tools
- Claude, Claude Code in VS Code, Lovable
- I built
- Claude skills for the team, and the decision wiki
How the team moved onto Claude
It took about a year, one habit at a time. It started with recording every call in July 2025. By November 2026 the team was answering questions from the decision wiki every day.
Every BA and design call gets recorded, so decisions don't live only in people's heads.
The team gets a company Claude plan. Analysts and designers compare AI prototypes before detailed design.
Analysts prototype in Claude instead of Figma Make. Designers start building in Lovable.
A shared library of the team's Claude skills. The team starts moving from Lovable to Claude Code.
New design work starts in a shared Claude workflow, with hands-on Claude Code training.
Every module gets an AI review before internal handoff, and serious findings get fixed first. The decision wiki goes live.
The decision wiki becomes the team's hub. Stories and designs get checked against it, and analysts answer QA and developer questions from it.
Two rules
- A person is always in the loop. AI suggests, and a person decides. That goes for specs, reviews, copy and every answer the team sends.
- Every AI finding gets an answer. It's either fixed or marked out of scope with a reason. Nothing gets quietly dropped.
The skills the team uses
A Claude skill is a set of written instructions Claude follows for one kind of job. These are some I built that the team uses every day.
- UX copy review. It reads the logic of a design, suggests the copy for each spot and strips out the AI-sounding filler. It gives the designer options to pick from, and the designer makes the call.
- Humanizer. It rewrites text so it's short, plain and reads like a person wrote it.
- Heuristic evaluation. It checks designs, prototypes, screenshots and UI descriptions against Nielsen's 10 usability heuristics, the 21 Laws of UX and cognitive load principles. I calibrated it for enterprise ERP and mapped it onto the product's design system.
- Mixed-language transcription. My team is multinational, and one call can switch between Urdu, Hindi and English. This skill transcribes video and audio on the team's own machines, tuned for exactly that mix.
- A wiki link for the handoff review skill. I added an extension that connects it to the decision wiki. When it reviews a design, it checks that every feature stakeholders approved is there and that nothing they didn't approve has crept in.
The team also runs skills for module reviews, user story checks and final QC before release.
Skills now work more like functions the team calls, so a big part of my job is knowing which one to call and when. Every analyst and designer has their own working rules for Claude, set up like mine, so a short prompt starts the right job. I trained the team on how to work with Claude, and I keep adding to the process as I find better ways of working.
More of the work
Tell me what's stuck
Send a few lines about your product, your team and the problem. If I'm not the right person for it, I'll say so.