One long Claude Code session per feature, or a fresh one for every step?

Fresh per step, with the "knows everything" part moved to files:

• CLAUDE.md: stable things (stack, conventions, commands).
• A plan file for the feature: steps, decisions, what's done.
• Git: what actually changed.

A fresh session reads those three in a minute and is sharper than a long one. The sloppiness you see is real: long contexts full of old attempts, dead ends and big file reads make the model pay attention to the wrong things.

I'm more in the middle. I keep a session through a step including its debugging, because the debugging history is useful ("we already ruled out X"). I clear when the step is committed.

Practical signal: when you notice yourself repeating a rule it followed earlier in the session, it's time to clear.

For refactors I go fresh per step, strictly. For bugs I keep the session until found, then clear before writing the fix, with a two line summary of the cause. The fix comes out cleaner without the hypotheses in the context.