I use Claude Code with a GeoDict skill myself. Still, I built the GeoDict Copilot as an application of its own, and a talk by Mitchell Troyanovsky, co-founder of Basis, describes exactly why.
In Why AI Makes You 100x Faster but Not Your Company he names two gaps. Coding agents push the complexity onto the user, who has to manage skills and memories and decide when to call the agent. And if every employee sets up the agent his own way, the work depends on that employee, while in a company the process is set from the top.
Working practice in the skills
These two gaps are the reason the Copilot is a desktop application. Its skills contain rules for working with GeoDict. One is a trial run before a study, another that every number in a report is computed by code from the result files. Every engineer in a team starts with the same 55 tools and 35 skills and describes the task in the chat.
GeoDict in an R&D team
In an R&D team, GeoDict studies can follow one working practice, whoever runs them, and no engineer has to set up and maintain an agent of their own first.