How I Actually Use Claude Code

I get the least value from Claude Code when I ask it for a snippet.

I get the most value when I give it a bounded outcome and let it work inside the repository: read the code, follow the existing patterns, make the change, run the checks, and show me the diff.

That is the shift: stop asking for code. Delegate a result.

Give it an outcome and a boundary

“Refactor auth” is not a task. It is a direction with no stopping point.

“Add refresh-token rotation to the existing auth flow, follow the pattern in src/auth/session.ts, and run the auth tests” is a task.

The agent knows what success looks like, where to begin, and how to verify the work.

Let the repository provide the context

Generic prompts produce generic code. I point Claude Code at the implementation it should imitate, then let it inspect the surrounding files.

Persistent rules belong in CLAUDE.md: commands, architecture decisions, coding standards, and the checks required before handoff. Task-specific context stays in the prompt.

The best work is usually boring and precise:

• Update a schema and every type that depends on it.
• Change a shared component without breaking its callers.
• Add edge-case tests to existing behavior.
• Apply the same error-handling pattern across several routes.

These are tedious enough to rush manually and structured enough to verify automatically.

Verification is part of the task

I do not ask Claude Code to “make the change.” I ask it to make the change, run the relevant tests, inspect the diff, and report anything it could not verify.

Then I read the diff.

Tests catch regressions. The diff catches bad assumptions. Neither replaces the other.

Reset when the goal gets blurry

When one task turns into three, I stop and open a fresh session. Long context feels productive while quietly weakening the original boundary.

I also use a separate review pass before the PR:

What could fail here? Which edge cases are missing? What would a senior reviewer push back on?

It catches real issues. It still does not know the business context, team history, or tradeoffs I left out.

Claude Code does not replace engineering judgment. It moves the mechanical work closer to the machine and leaves the judgment visible in the diff.