Sep 3, 2026 · by Terry Peterson · claudeworkflowhabitscli
Who this is for: People who use Claude Code in the terminal (CLI). Browser-only AI users: the idea carries over, but use your AI to re-write the steps.
AI sessions end. Context windows fill up, laptops close, the next morning arrives. The
AI session that helped you yesterday remembers none of it. For a one-off question that’s fine. For a project that runs for months, it’s the whole problem.
The fix I landed on was to have AI write it’s future self a set of notes, and the first thing the next session does is read it.
I use to manually request for Claude to write itself a handoff each time, but the results weren’t always consistent. Since I run multiple projects, I needed to standardize this, so I had Claude Code create a /handoff skill. This way all my project know to store a
session-latest.md in the repo, and it always answers four
questions:
What is the state right now? Is anything half-done, or unpushed, or running in the background, and includes instructions to check, not just assume from memory.
What did this session do, summarized in plain English, not commit messages.
What must be verified next time? Code that passed a test today is not the same as verifying an overnight job produces the right results tomorrow. The next session is the one that will see the result, so tell it where to look.
What comes next, in order, plus the decisions that were already made so they do not get brought-up or asked again.
A few things that turned out to matter A TON:
One file, overwritten every time. Git history is the archive. A folder of dated
handoffs just becomes a folder nobody reads.
Write it for a reader with less context than the writer. The session that writes it
knows everything; the session that reads it knows nothing. Names of files and their repository, not high-level descriptions of them like “check the logs.”
Include “do not relitigate”. AI is eager to re-open settled questions. A short
list of decisions and why they were made saves a surprising amount of time (and tokens).
Make the AI write it, not you. It has the full context of what it just did. Your job
is to read it and correct anything that’s wrong. Admittedly with a dozen on-going Claude Code projects, I no longer read these. But early on I had to ask Claude to refine the skill a few times based on what I found.
If you only adopt one habit from this blog, this is the one.