Agentic Coding #5 — The Coding Workflow

The Workflow with Claude Code

The actions are what gets handed directly to Claude Code. You tell Claude: Read through this project. Claude then finds the associated implementation steps, devised by the user, and knows what it needs to work through.

There is a CODING.md that describes how Claude processes each step. The procedure is the same for every action:

First, a plan is created. For this, there are three plan agents that examine the action from different perspectives and each produce a partial plan. The main agent then consolidates these results into an overall plan. This plan is presented to the user, and implementation only begins once it has been approved.

During implementation, tests are always implemented alongside the code. Afterward comes a review process: Three code review agents examine the result, and three test agents examine the tests. The main agent summarizes the findings from both the code review and the test agents and implements the insights.

This is how actions are processed one by one. When all actions of a project have been completed — if there were multiple — a code review and a test review are run across the entire project. This ensures that not only the individual steps but also their interplay is verified.

Guardrails and Best Practices

The CODING.md contains not only the workflow but also clear instructions that keep Claude in check. There is a multi-eye principle through the reviews. It specifies that the linter must be used. And above all, there are rules against Claude's tendency to take shortcuts or make assumptions.

Claude must not make assumptions but must always check the source code. Claude must adhere to the defined work steps. If Claude considers something completely unnecessary, it may not simply omit it but must at least ask whether it may be omitted — with a justification. That is the point: It is often correct that something appears superfluous, but the decision lies with the user, not with Claude.