Handing work from one Claude Code session to the next
A long Claude Code run ends, and everything it worked out goes with it: why that approach was rejected, what the constraint was, what you told it in the fourth message. The next session starts by reading the code and guessing at the rest.
Write the decisions and their reasons into a branch of your outline as you go, and the next session reads that branch instead of re-deriving everything from the code.
Set it up
- In a terminal: claude mcp add --transport http --scope user pando https://pando.ink/mcp
- Start claude, run /mcp, select pando and choose Authenticate.
What a handoff actually needs
Not a transcript. A transcript is the thing the next session would have to read in full to learn three facts, and it will not.
- The decisions, each with the reason underneath it.
- What was tried and rejected, so it is not tried again.
- The constraint that is not visible in the code.
- What is still open, as a question rather than as a summary.
Reading it back without wasting the first turn
The recall tool answers a question rather than dumping a branch, and the changes tool answers with only what moved since a cursor you hold. Measured on our own outline: a changes call with the last cursor answered 1,060 bytes where the same tree cost 962,608. A session that starts by reading everything has spent its first turn on bytes nobody asked for.
Why not compaction
Compaction shortens a conversation and keeps it inside that conversation. This is the other axis: the branch is outside every session, so a second machine, a phone, or a different agent entirely reads the same thing. And you can read it too, and correct the line that went wrong.
Connect an agent, about two minutes · The twelve tools · Which note apps an agent can reach · Who runs this