How to Use Claude Code for Real Development Tasks

Claude Code is a separate agentic coding tool from Anthropic that works directly in your development environment. Unlike chat-based Claude, it can read your codebase, run commands, and make file changes — which makes scope control even more important. It is available on eligible plans; check current Anthropic documentation for availability.

What Claude Code can do

• Read and navigate your codebase to understand existing patterns
• Write and edit code across multiple files in a coordinated way
• Run tests, builds, and linters to verify changes
• Create commits with descriptive messages
• Explain code and architecture when you ask before making changes

The plan-before-edit workflow

The most common mistake with Claude Code is letting it start editing immediately. Ask for a plan first. Review the plan. Then approve the implementation.

I want to add rate limiting to our API endpoints.
Before writing any code:
1. Read the existing middleware in src/middleware/
2. List which endpoints currently have no rate limiting
3. Propose an approach that uses our existing Redis connection
4. Estimate which files will change

Do not make any changes until I review the plan.

Scoping tasks effectively

Claude Code works best on tasks with clear boundaries. 'Refactor the entire backend' is too broad. 'Add input validation to the three /api/billing endpoints' is well-scoped.

Bad scope

Clean up the codebase and fix any issues you find.

Better scope

Add Zod validation to POST /api/billing/subscribe.
Match the validation pattern in POST /api/auth/register.
Only touch src/app/api/billing/subscribe/route.ts and its test.
Run the existing tests after your changes to confirm nothing breaks.

Review checkpoints

For multi-step tasks, break the work into checkpoints. Review at each checkpoint before proceeding.

• Checkpoint 1: Plan — review the approach and affected files
• Checkpoint 2: Core change — review the main implementation
• Checkpoint 3: Tests — review test coverage and edge cases
• Checkpoint 4: Cleanup — review for leftover debug code or TODOs

When to step in

• When Claude Code suggests changing files outside the agreed scope
• When the plan involves installing new dependencies you have not vetted
• When the approach diverges from your existing patterns without a good reason
• When tests pass but you do not understand why the fix works

Practical task examples

• Add a migration and update the Prisma schema for a new field
• Write integration tests for an existing endpoint that has none
• Refactor a component to extract a reusable hook, keeping the same behavior
• Update import paths after moving a module to a new directory

Common Claude Code mistakes

• Giving broad, open-ended tasks with no scope boundary
• Skipping the plan step and letting Claude Code make changes immediately
• Not reviewing the diff before committing — always read what changed
• Assuming tests passing means the change is correct — tests can be incomplete

Related reading: /blog/how-to-use-claude-for-coding for chat-based coding and /blog/how-claude-agents-work for how agentic workflows connect to tools like Claude Code.