Why a fresh chat beats a long one
Two reasons — cost and quality — and they both go the same direction.
Not filmed yet
The lesson is written out below. The video comes after.
It feels wasteful to close a session that “knows” what you are doing. It is not, for two separate reasons.
The first is cost. Every message you send re-reads the entire conversation up to that point, so a fat chat multiplies your spend on every single turn. The tenth message in a long session costs many times what it costs in a short one, for the same question.
The second is quality, and it matters more. Models get fuzzier on material buried deep in a long context, so the thing you established two hours ago is exactly the thing most likely to get misremembered. A fresh session pointed at a file containing the same decision is reading it at full strength.
Which means the discipline is not “keep the chat alive.” It is “get the decision out of the chat and into a file, then start fresh.”
The one thing to keep
A fresh chat reading good files gets better information, not worse.
More in Claude Code
Next to this one
Skills — teach it once and it knows forever
Do the task, get it right, then say "turn what we just did into a skill." The work becomes the training material.
Claude CodeHooks — rules that fire without asking
Automatic tripwires, so the discipline is not left to your memory.
Claude CodePlan mode — make it think before it builds
A read-only mode where it can study and propose but cannot touch anything.