Feature Workflow

No more babysitting the AI

Working with a coding assistant often means answering a question every thirty seconds. Approve this plan, create that file, accept this approach. You end up supervising the tool instead of getting on with things.

Feature Workflow is a plugin for Claude Code, Anthropic's coding agent. One command describes the feature you want, and the rest runs on its own, through to tested and reviewed code.

Eight stages, each handed to specialists

It starts with analysis: several agents explore the codebase in parallel to work out its stack, habits and conventions. Then planning, where three competing plans are drafted (simplest, cleanest architecture, most pragmatic) before one is picked.

Next comes implementation by teams of agents working side by side on the interface, the server and the database, followed by strict type and style checks. Three specialised reviewers go over the result and fix what they find. Tests are detected, written and run automatically, and the documentation follows.

Language-specific best practice is fetched at the time of the work, rather than relying on whatever the model happened to learn.

Scaled to the size of the job

A small change doesn't need a whole team of agents, so a lighter mode does the same work with far fewer resources. For a vague request, on the other hand, an opening questionnaire helps pin the feature down before anything starts. Progress and the list of finished features can be checked at any time.

What it says about how we work

Feature Workflow formalises a method we also follow by hand: understand the existing code before touching it, compare several approaches, review, test, document. That is the discipline we bring to client projects, with or without AI.