Resume the project without rereading every AI chat
The goal is not to remember the conversation. It is to recover the last verified project state and choose the next bounded action.
Read the roadmap before opening a new chat
Locate the active step, its completion criteria, and any blocked dependency. This restores project intent before a new Agent starts proposing work.
Separate claims from verified facts
Review the last result, touched files, test output, and handoff. Treat unverified Agent summaries as leads, not as completed work.
Check the current repository state
Code and logs may have changed since the last session. Confirm the working tree and run the narrowest relevant validation before continuing.
Continue with one explicit outcome
Tell the Agent what must change, what must remain stable, and how completion will be checked. Do not ask it to ‘continue’ without a boundary.
Leave a better handoff than you found
Record the result, remaining risk, and next action. The next session should start from project facts instead of chat archaeology.