After six weeks of daily Claude Code use, here are the five mistakes I made most consistently and what I replaced them with.
Mistake 1: Vague task descriptions
What I did: "Fix the login bug."
What I do now: "The login endpoint returns 500 when the email field is missing from the request body. Fix it to return 400 with a message explaining which fields are required. Regression test: add a test that confirms missing email returns 400."
Specific tasks produce specific results. Vague tasks produce Claude guessing at what you meant.
Mistake 2: Not reading diffs
What I did: glance at the diff, note the main change, commit.
What I do now: read every changed file end to end before committing. Claude made a small change to a config file I wasn't watching. It was correct but unexpected — I want to know about it before it ships.
Mistake 3: Trusting library API knowledge
What I did: ask Claude to implement with a library, trust the output.
What I do now: check the version I'm actually using, verify key API calls against the docs for that version. Claude's library knowledge can be one or two major versions behind.
Mistake 4: Sessions too long
What I did: start a session, work all afternoon in it, keep adding tasks.
What I do now: end the session when the task is done. New task = new session. Fresh context means Claude isn't carrying forward assumptions from three tasks ago.
Mistake 5: No CLAUDE.md for the first week
What I did: skip writing CLAUDE.md because it seemed like overhead.
What I do now: write CLAUDE.md before any feature work. 10 lines minimum: build commands, naming conventions, what not to touch. Quality went up noticeably once Claude had the project context.
All five come down to the same thing: Claude performs better with clear constraints and fresh context.
Top comments (0)