Skills and AI
The code is half of what you receive, and the other half is how to change it.
A skill is a written procedure an AI assistant follows to do one job the way this codebase does it.
There are skills for scaffolding a new command slice, adding an account type by extending an existing one, writing a component test the way the suite expects, writing a help article and syncing settings, 38 of them today, and they are how we build. The list, by job → They ship with the source as plain text, so your team extends the engine the way we do from day one, and changes the procedures, adds its own, or adapts the loop to its own environment the moment it wants to.
The documentation is kept current by the same process that ships the code.
A decision record is written before a change, and when it ships it is marked with what actually happened. A help article goes out in the same commit as the screen it explains, and the domain model and the capability atlas are generated from the code. Most vendors hand over a wiki that was true at go-live.
- Read the designthe decision records, in order, with the reasons
- Run the looprecord, grill, plan, bites, gates, with the same skills
- Add a productan account type from a template, tested with the forecaster
- Add a capabilitya new slice behind the contracts boundary, gated like ours
- Take the update stream, or nota release is a snapshot, and taking our ongoing updates is your choice
Vendor risk becomes a choice you make rather than a risk you carry. We're a two-person vendor and we know how that sounds, but it stops being a bus-factor problem once you hold the code, the tools and the history.