AI work method
When Claude can lose track of context: what to write down before a long session
It was following the plan an hour ago. Now it suggests the approach you dropped this morning.
In Claude Code, each new session starts with a fresh context window, without the conversation history from previous sessions. As a session fills up, Claude compacts it into a summary, and Anthropic says detailed instructions from early in the conversation may be lost. In Claude chat, memory may carry some context into a new chat when it is on. Standing rules belong in a notes file such as CLAUDE.md; what was decided and why belongs in a decision log you keep.
Sources, checked 2026-11-02: How Claude Code works; How Claude remembers your project; Use Claude’s chat search and memory to build on previous context. Features change, so check these pages for the current wording.
Two situations
A long session. Signs: in my experience, it starts going against something agreed early on, or asks about a choice you already made. What to do: write the decision down as soon as it is made, so you can paste it back.
A new session. Signs: it starts with no idea of yesterday’s work. What to do: open with a short state note and the decisions that still hold.
What to record, and when
Record each decision as you make it, not at the end of the day. By the end of a long session, you may no longer agree with the AI on what was decided.
A decision-log entry you can copy
Date:
Decision:
Alternatives considered:
Reason:
Replaces:Illustrative example (one week of a made-up writing project):
Date: Monday
Decision: The guide is written for first-time landlords, not investors.
Alternatives considered: both audiences; investors only.
Reason: the interview notes are all from first-time landlords.
Replaces: nothing (first entry).
Date: Thursday
Decision: Chapter 3 drops the tax section.
Alternatives considered: a short summary; a link out.
Reason: it needs an accountant's review I don't have.
Replaces: Monday's outline, item 3.A short state note for the next session
Start the next session with the goal, the open question and the next step, plus the decision log. For the full version, use a seven-line handoff note.
My tool has a notes file. Isn’t that enough?
Anthropic’s docs say Claude treats CLAUDE.md as context, not enforced configuration, and that Claude reads it and tries to follow it without a guarantee of strict compliance (How Claude remembers your project, checked 2026-11-02). A notes file holds standing rules. A decision log does a different job: it records what was decided, when, and why.
What this does not fix
The AI can still make mistakes with a good log in front of it. Check the output. How Mothex grew, one failure at a time is the history of the method; Mothex is the written method these notes come from.
Get the free Handoff Kit (six plain-text files, no email)
Claude is Anthropic’s product. Lizvo is independent and not affiliated with Anthropic.
— Lizvo
General information only, not professional advice. See the disclaimers.