How does Kliwe analyse a chess game?

The engine evaluates every position in the game, the largest swings are picked out, and DeepSeek turns those moments into an explanation you can read.

Yes. The evaluation is fast rather than deep, and the write-up is a coach's opinion, not a proof. That is exactly why every analysis opens a thread you can argue in.

Is my game published?

Games you save are public so the community can discuss them. Games you do not save are never stored.

How AI chess analysis works on Kliwe

The short version

The engine does the counting. The language model does the explaining. Neither is asked to do the other's job.

Step by step

1. The game is stored with a position for every move. Not just the notation: the full board state after each half-move, so nothing has to be recomputed later and nothing can drift.

2. The engine evaluates every position. This is a fast material-and-placement evaluation, in pawns, positive for White. It is not a deep search — it is a measuring tape, not a grandmaster.

3. The swings are pulled out. Where the evaluation moves sharply from one move to the next, something happened: a piece was hung, a tactic landed, a plan collapsed. Those moments are the skeleton of the write-up.

4. DeepSeek writes it up. The model is given the move list, the result, the opening and those swings, and is told to explain the game to a club player. It is instructed to discuss only moves that actually appear in the game, and the moves it cites are checked against the real move list before anything is published. A comment on a move that was never played is dropped.

5. It becomes a thread. The analysis is posted to the community with a question at the end, because the interesting part is the disagreement.

What this is good at

Naming what went wrong in words you can use in your next game.
Finding the moment a game turned, which is usually earlier than you remember.
Explaining an opening you have played a hundred times without reading about.

What it is not good at

Precise evaluations. The engine's number is a hint, not a verdict. If it says +1.4 you should read "White is better", not "White is 1.4 pawns better".
Deep tactics. A fast evaluation misses long forcing lines. If the analysis says a move was a blunder and you can show it was not, you are probably right, and the thread is the place to say so.
Certainty. A language model writes fluently whether or not it is correct. That is why the analysis is published next to a reply box rather than on its own.

Your games

Games you finish on the board are saved with their own page and analysed. Games you abandon are not stored. Saved games are public — that is the point of the archive — so play accordingly.