ChessCoach: A New Chess Engine with its Own Commentary

  • Post author:
  • Post category:General
  • Reading time:9 mins read

A new and unique chess engine was released in September called ChessCoach. ChessCoach is a very strong engine, at about 3450 Elo and will of course beat even the strongest GMs.

What makes ChessCoach unique is that it adds commentary to its own games. This can be a groundbreaking advancement for chess improvers everywhere. Imagine analyzing your game with a ChessCoach engine and having the engine explain to you, in English where you went wrong and why.

The ChessCoach engine used academic papers from Carnegie Mellon and Peking University to teach an AI to generate natural language processing to analyze games.

You can challenge the engine on Lichess. Its username is PlayChessCoach. As of this writing, ChessCoach is rated roughly 2500 across all bullet, blitz and rapid time controls.

Let’s try it

I challenged the engine to a game to see what kind of commentary it offers. I was lucky enough to have the white pieces, and ChessCoach has black.

The Opening

We got a symmetrical queen’s pawn game.

4…h6 is an interesting choice. A bit premature to give a square to retreat the bishop, but the move is just fine. Let’s go over what ChessCoach says after these initial moves

  1. d4 Nf6

ChessCoach: “This is a standard response to 1.d4. It is the most common response to 1.d4 (the c pawn is still protected by the c pawn).”

The engine is correct that Nf6 is a standard response to d4, and it is indeed most common. The c pawn being protected by the c-pawn can is debatable.

2. Nf3 d5 3. g3 Bf5
“Black wants to develop his bishop on the long diagonal, but the knight on f3 is not so bad as it looks.”

This is a very strange comment. Black isn’t doing anything to develop their bishop to the long diagonal. White is, however, with Bg2 to come. The knight of f3 doesn’t look bad- at least to my human eyes.

4. Bg2 h6 (Would have been nice to see some commentary here since it’s an offbeat opening choice)

chesscoach opening

5. O-O e6
“The idea is to play …c6 and exchange the dark-squared bishops, which will give White the chance to play d4 at some point.”

🤔 This is complete nonsense. c6 is a fine move (and it does get played next), but has nothing to do with exchanging dark squared bishops. And white already has a pawn on d4. Let’s continue on.

7. c4 c6 8. Nc3 Be7
“The bishop is well placed on e7 and the knight on f6 is very well placed.”

Very nice! Agreed!


I think we can call this the middle game. We’re already in rare-ish opening territory and the upcoming moves are not standard theory. We are now at the following position. I play Qb3 here to pressure the b7 pawn.

ChessCoach middlegame

8. Qb3 Qb6 9. c5 Qxb3
“I was surprised to see black exchange. But, it is not a good idea to exchange when behind.”

I kind of agree that black shouldn’t necessarily exchange here. We’re in a typical position where either side capturing the queen will give the other side a half open a-file to work with. I also agree that it isn’t a good idea to “exchange when behind” but black isn’t “behind” here, in any sense of the word.

10. axb3 a6 11. b4

Middlegame after queens were traded

11… O-O 12. b5
A typical idea in this type of positions. White is trying to get rid of the weakness of the b5-pawn, but Black is doing fine.”

This is my favorite comment so far! It’s very close. b4 is definitely a very typical idea in these types of positions. You will also see these types of positions come out of the London System. See how NM Matt Jensen handled it in a video on our YouTube channel. My only nitpick is that white isn’t trying to get rid of the weakness of the b5 pawn, but rather white is trying to create a weakness in black’s camp. Black is still doing fine, as black is an engine and plays perfectly.

12…cxd5 13. Nxb5 (exploiting the a-file pin) Nc6 14. Nc3 (retreating the knight since the a8 rook is now defended) Nd7
“The most logical continuation, as now White has to decide whether to play d2-d3 or e2-e4.”

This is again nonsense garbage. d2-d3 is not a possible move and e2-e4, while legal, just blunders a pawn.

15. Rd1 b6 16. cxb6
“[cxb6] is a typical exchange sacrifice in this structure.”

This is not an exchange sacrifice.

Middlegame after cxb6

16… Nxf6 17. b3 (keeping the knight out of c4) Rfc1
“The b3-pawn is not hanging, so White’s pieces are uncoordinated.”

The b3 pawn is not defended, but nothing is attacking it, thus it isn’t hanging, very nice! I also cannot argue that white’s pieces aren’t coordinated, but I’ve seen worse coordination!

18. Bb2 Bc2 19. Rb1 (mouse slipping the rook to b1 instead of c1) Bxb1
“White’s knight is still pinned, but the pawn is now pinned. Black’s knight is attacking the bishop on b2.”

This is hilarious. None of this is true at all. Swing and a miss.

The game went on a a few more moves, but I’ll leave it here.


The new ChessCoach engine is very strong. It had a winning position against me in just 18 moves, after I felt I did nothing too silly. The commentary did have very good chess lingo, using phrases like “exchange sacrifice” “typical idea in this structure” and talking about “hanging pieces” as well as and other tactical motifs. The language does feel very natural and not computer-like at all.

The actual commentary, however, did not hold much chess truth. It would be groundbreaking if this engine could evolve to give insights into mistakes and blunders over the course of a game. Chess players everywhere would love to analyze a game with this engine to get ideas and pinpoint their mistakes. It looks like we have a ways to go before that is possible.