When: a long conversation needs summarising and you want the summary to be right, not just plausible.
Built from: the transcript, agents, the context window, re-alignment.
The problem: if you ask a conversation to summarise itself at the end, it summarises from a full window, seen from the end. It will do its best, but at that point it is unsure what was fact, what was corrected, and what was dropped along the way.
The fix: summarise from the record, not from memory.
- Save the session transcript to a temporary file. In Claude Code the transcript is kept as a JSONL file. This is also why you keep your logs past the default: the transcript is the record.
- Hand it to an agent. An agent is a fresh context window with one job, not polluted by the session it is reading about. Tell it to read the transcript and weigh what the person said over what the AI said, and to treat later messages as corrections to earlier ones, since the last word on a point is the one that stood.
- The agent writes its view of what was decided and what is still open.
- Bring that view back into the original conversation and hold it against what the conversation itself believes. Anything that doesn't match gets surfaced to the person, not resolved quietly.
As a skill: ask the AI to turn this page into a skill. It already knows where the transcript lives.