Review method · 01

How to review a Go game with AI

A four-step routine that does not chase every number: reconstruct your thinking, use KataGo to find turning points, and keep lessons you can use next game.

1. Replay once before opening the AI

   Move quickly through the game and mark where you thought for a long time, felt unsure, or were surprised by the result. Write down what worried you then, so you can compare thought processes rather than merely copy an answer.

   On the first replay, do not recalculate every variation. Mark only three kinds of moment: where you slowed down, where you regretted the move immediately, and where the post-game result surprised you. Beside each mark, write the question you had before AI changed your memory, such as “Can this group live?” or “Must I keep sente here?” Those untouched questions later reveal whether the real training need is reading, direction, or risk judgment.

2. Find turning points with both graphs

   Import the SGF and start analysis. Scan the overall trend, then choose two or three positions with a meaningful change in win rate or score lead and enough context to understand it. Late fluctuations in an already decided game are rarely the best place to begin.

   The two graphs answer different questions. Win rate estimates the chance of winning if both sides continue strongly; score lead estimates the board margin. A sharp win-rate swing with a small score change often means the game was already extremely close. When both graphs move together, the position usually deserves attention. Place your marker on the move before the swing so you study the position in which the decision was actually made.

3. Compare branches in TreeView

   Start with the game branch, then open KataGo candidates in TreeView. Return to the same fork each time and ask what threat a branch answers, what sente it preserves, and which group becomes lighter or heavier.

   A candidate list is not an answer sheet. Keep the game move L11 in TreeView, then open the J11 and N7 candidate branches and identify what each changes in move order, group safety, or sente. Return to move 77 and create a branch for your most natural reply. TreeView preserves the historical line while you switch among alternatives, so the actual game never becomes confused with an experiment. The result is a reusable reason rather than an isolated coordinate.

4. Rewrite the result as a future cue

   Do not copy a long variation. Give each key position one action you can recall, such as “check my weakest group before attacking” or “when ahead, compare score lead instead of watching win rate alone.”

   Keep only one to three notes from a game. A useful format is “position signal → old habit → next action,” for example: “They have a weak group but mine is unsettled → attack immediately → first locate my weakest group.” Avoid saving coordinates and twenty-move lines unless you are studying a joseki that truly requires recall. Reopen only the position a few days later; if you can still name the check to perform, the note is useful.