AI coding assistant
An AI coding assistant adapted to the team's way of working
A coding assistant that starts from the team's context, with specialized roles, written conventions and human verification before changes land.
- The problem
- A generic assistant doesn't know a codebase spread across many repositories or the team's process.
- The decision
- Specialized roles, conventions as editable documents and human verification before changes land.
- How it ended
- An internal tool that follows the workflow the team already had.
The problem
The platform is spread across many repositories. A generic coding assistant doesn't know how they relate, how the team writes a task, or how a change gets checked before it goes to testing. I wanted a tool that started from that context: one that helps implement and review code while following the process the team already uses.
Decisions
Not depending on a single model provider. I kept the assistant's logic separate from the model that runs it, so the provider can change without rewriting the tool. Cost: differences between providers have to be handled by hand.
Specialized roles instead of one assistant for everything. I defined separate roles, for example one that implements and one that only reviews, each with access limited to what it needs. Cost: more pieces to configure and keep consistent with each other.
Conventions written as editable documents. The team's way of working lives in documents anyone can read and change, instead of being hidden in code. Cost: behavior lives in text, and a change can only be checked by running it.
A person verifies before a change lands. The assistant proposes and tests, but someone on the team makes the final call. Cost: the flow still has a manual step and moves more slowly.
Outcome
- An internal tool that knows the structure of the code and the team's process.
- Less time spent explaining context to the assistant on each task.
- Changes that are reviewed by a person before they land.
What I'd do differently
I'd keep a record of the tool's changes from the start and ship them in smaller, more spaced-out steps, instead of piling a lot of work up at the end.