In short The core working rhythm of Claude Code: not "ask it to write code and hope," but four steps — explore → plan → code → commit. Each step cuts the risk of getting the wrong thing.
Why a rhythm, not "just ask"
The beginner's temptation: "write me feature X" — and accept the result at once. On a real codebase that turns to mush: the agent lacks context, works in the wrong style, breaks the neighbouring code. EPCC forces you to understand and agree first, then do — which is how a good human engineer works too.
The four steps
1. Explore
Claude reads the relevant code: which files, how it's built, the conventions, where the edges are. It changes nothing yet. The goal is to gather context so the solution fits the existing project, not sits beside it. It helps to limit what to read so context doesn't bloat (agent context management).
2. Plan
From what it explored, Claude proposes a plan: what to change and how, in what steps. You review and fix it BEFORE any code is written. This is the cheapest moment to catch a mistake — a plan is easy to rewrite, finished code is expensive to redo. In Claude Code this is what Plan Mode is for.
3. Code
Only now does Claude write code to the approved plan — editing files, running commands, running tests, fixing as it goes. You watch and step in if it drifts.
4. Commit
The changes are reviewed and committed. Before committing — code review: read the diff with your own eyes, and it helps to have a "fresh" reviewer with clean context too (a separate session, not blurred by the work). A clean commit means clear history and the ability to roll back.
Analogy (a renovation): survey and measure (explore) → agree the estimate and plan (plan) → do the work (code) → accept and sign off (commit). Knocking down walls with no plan is a recipe for disaster.
Why it works
Mistakes are caught at the plan, where they're cheap to fix; the code fits the project's real context (the explore step); a human in the loop at plan and review — you approve the intent and the result, while the agent does the grind.
Related
The source is the Claude Code 101 course. Related: Claude Code — the terminal agent, Claude Code setup, agent context management, prompting and AI Fluency.