Keep the project’s decisions in a file every agent reads, ask the agent for a short handover before you close a long chat, then start a new session from that handover. In our own use (39 handovers), starting fresh from a handover cut the context tokens sent per call by a median of about 70% (about 55% by cost). Method below.
Why do long AI coding chats use more tokens?
Each time a coding agent calls its model, it sends the conversation so far along with your new message. Early in a chat that’s small. A few hours in, every call carries everything: each file it read, each test it ran, each wrong turn. The calls get bigger, your plan’s limits fill faster, and the instructions you gave at the start get summarised away.
Starting a new session fixes the size at once. People put it off because a new session knows nothing, and explaining the project again costs time and tokens of its own. The answer is to keep the project’s context somewhere other than the chat.
What we measured
In our own use, across 39 handovers, starting a fresh session from a handover cut the context tokens sent per call by a median of about 70% over the 20 calls after it, against the 20 before (about 55% by cost).
- From 658 Claude Code sessions saved on the founder’s own computer; the 39 handovers fall between 8 and 21 Oct 2026.
- A handover: a fresh session whose first prompt is a handover, started within 6 hours of a session of 100+ calls in the same repository.
- Median context tokens sent per call: the first 20 calls after the handover vs the last 20 before it.
- Cost weighting: cache reads 0.1×, cache writes 1.25× (1-hour writes 2×), output 5×.
- One developer (the founder), mostly large-context models. Results vary.
That’s what the freedom to start fresh was worth in one developer’s work, not a promise for yours. A fresh session has less context to send with each call; we didn’t test other explanations.
1. Keep decisions in the project, not the chat
A new session can only pick up what’s written down. Keep what was decided, and why, in a file the repository tracks, and make every agent read it on start. Then a fresh session already knows the rules. The decision log guide has the format.
2. Write a handover before you close
Ask the agent to write down where things stand, then read it yourself. A handover is short: what’s done, what’s next, and what’s still open.
Point to the decisions rather than copying them in, so the handover can’t go out of date when a decision changes.
3. Start a new session from it
In Claude Code, /clear empties the conversation and keeps you in the same folder. With any agent you can also close it and start it again. Then give it one instruction:
Two Claude Code commands look similar and do something different. claude --continue and /resume reopen an earlier conversation as it was, so its size comes back with it. /compact shrinks the current conversation into a summary the agent writes. A handover is the short version you’ve read and checked.
4. Know when to do it
- When a job is finished and the next one is different.
- When a chat has run for hours, or the agent starts repeating itself.
- Before you hand the work to a different agent, so it starts from the same page. Running several at once? Give each its own worktree.
Long chats are also where agents drift from what you asked: see stopping agents drifting.