Same tool. Three projects. Three completely different approaches to Claude Code

Claude Code is not a one-size-fits-all tool. How you start matters more than how much you use it. Three project types. Three completely different approaches.

Greenfield — start with CLAUDE.md, not code

You have the cleanest canvas you'll ever get. Use Claude Code to scaffold structure, tooling, and CI first. But before writing a single line of app code — write your CLAUDE.md. It's the file Claude Code reads every session. Without it, you're re-explaining your stack and conventions from scratch every single time.

Brownfield — generate CLAUDE.md first, that's your understanding

Don't just jump in. And don't write the CLAUDE.md manually either. Ask Claude Code to explore the repo and produce it — read the structure, the routes, the services, the config. Reviewing what it generates is how both you and Claude Code build a shared mental model before anything gets changed. That's your "understand first" step.

One more thing: be explicit about scope. Claude Code will helpfully refactor things it passes through. On a live system, a PR touching 40 files when you needed 4 is a real problem. Tell it upfront — "only touch what this story requires."

Tech modernization — every difference is a bug

Same functionality, new stack. Claude Code is great at reading legacy code and understanding intent. But it will silently improve things during translation — fix a bug that was actually a feature, change an error message another system depends on. Our rule: any behavioral difference between old and new is a defect, not an upgrade. We encode that in CLAUDE.md and enforce it with parity tests in CI. Pipeline fails if outputs diverge. No exceptions.

Below three things that apply across all of these scenarios that most teams figure out too late.

Plan before you code. Claude Code has a Plan Mode (Shift+Tab twice). Use it. For anything non-trivial, ask Claude Code to propose a plan and show it to you before writing a single file. Review it. Push back. Correct the direction. A five-minute planning conversation saves hours of wrong-direction implementation.

Context degrades over a long session. The longer a Claude Code session runs, the more the quality quietly drops. Old instructions start getting ignored. Code patterns drift. The session is carrying too much history. The fix is simple: use the /clear command between unrelated tasks. Start fresh. It's not a loss of progress — it's a quality reset.

Commit constantly. Claude Code can touch a lot of files quickly. If something goes sideways 45 minutes into a session and you have no checkpoint, you're doing a painful manual review of every file it touched. We ask Claude Code to commit after every logical step — route, service, repository, tests — as separate commits. Each commit must leave the tests green. This gives you a clean rollback at every point.

The teams getting the most out of Claude Code aren't the ones using it the most. They're the ones using it with the most intention — clear context, defined scope, and human judgment at the decision points that matter. Claude Code is genuinely excellent at the implementation. The architecture, the tradeoffs, the risk assessment — that part is still yours.

#ClaudeCode#EngineeringLeadership#SoftwareDevelopment#AITools#TechModernization#DeveloperProductivity