Why
- The model takes everything into account. A backup question asked after an hour on Traefik gets an answer shaped by Traefik: examples with its paths, the same level of detail, the same assumptions.
- Every turn costs more. The history rides along in every request. The third task in a chat pays for the input of the first two — both in the API and in subscription limits (Tokens).
- Mistakes stay. A wrong answer from the first task is a template for the model too (Edit, don’t argue).
What it looks like
| One chat before | After |
|---|---|
| “why the 502”, then “rewrite the compose file”, then “and the backup while you’re at it”, then “an email to the ISP” | four chats; the second gets the first one’s conclusion: “the 502 is caused by the service port in a label; I need a compose file with the fix” |
In Claude Code it’s /clear before a new task. Project rules aren’t lost: they live in CLAUDE.md, not in the history
(Long conversations).
When to stay in the old chat
If the next question continues the same task and needs details from the history: “now the same for the second server”. The sign it’s time to leave: the new question can be asked without a single reference to the previous one.