The marketing shows the good parts. Here's what you figure out through experience.
The first week you'll be slower
Every new tool has an adjustment cost. With Claude Code it's learning which tasks to delegate, how to write prompts, and how much to trust the output. Your throughput will drop before it rises. Most people either push through or give up here.
Claude will confidently do the wrong thing
Claude doesn't say "I'm not sure" often enough. It picks an approach and executes with high confidence. Sometimes the approach is wrong and the execution is impressive. The only protection is reading diffs carefully before committing.
Your CLAUDE.md will be wrong the first time
Every CLAUDE.md starts as a guess. You discover what needs to be in it by running sessions and noticing what Claude gets wrong. The file improves over weeks, not days. The first version is just a starting point.
Session restarts lose context
If your session dies mid-task, Claude doesn't remember what it was doing. Without a state file, you're back to explaining the task from scratch. The state file pattern (tasks/current-task.md, updated after each step) solves this. Set it up before you need it.
The longer the session, the worse the quality
Context fills as the session progresses. A Claude session working for 3 hours on a complex codebase is carrying a lot of accumulated context. Quality tends to drift in long sessions. Starting fresh more often, with clear context passing between sessions, produces better results than one marathon session.
You'll still write some code yourself
Claude Code is a tool, not a replacement. The parts that require your judgment — the design decisions, the business logic that depends on context only you have, the security-critical code you need to own — those stay with you. The best mental model is: Claude handles the what-you-already-know-how-to-do. You handle the what-requires-thinking.
Top comments (0)