The ChatGPT chess coach prompt (and 7 more that help you improve)

A good ChatGPT chess coach prompt doesn't ask the model to calculate. It gives the model the jobs it does well, such as explaining ideas, planning your study and asking you good questions, and it leaves the moves to a chess engine. Below are eight copy-paste prompts for improving at chess with an AI assistant: setting up the coach, a study plan, explaining an engine line, studying an opening, reflecting after a game, puzzle-style training, learning a concept and finding patterns in your mistakes. They work in ChatGPT, Claude or any other assistant.

Replace everything in [square brackets] with your own details before you send a prompt.

The one rule: the engine finds moves, the AI explains them

Language models are not chess engines. In a study posted in August 2026, GPT-5.4 without tools got 22.0% of the factual claims in its chess commentary wrong, and the authors found that giving models tools improved factual accuracy substantially (Hebbar et al., arXiv). A Chess.com blogger who used ChatGPT 5.5 as a coach in May 2026 found it useful on plans and ideas but unreliable on concrete moves (Chess.com blog). We look at the details in Can ChatGPT analyze chess games?

So, whatever prompt you use, never trust the model on these four things:

• Evaluations. "White is clearly better" or "+1.5" from a chatbot is a guess unless you pasted that number from an engine.
• Long variations. Every extra move is another chance for the model to lose track of a piece. Treat any line longer than three or four moves as unchecked.
• Legality. Models still try impossible moves. On one public leaderboard, even the strongest models from OpenAI and Anthropic try an illegal move on up to about 4% of their moves, and chat models without extended reasoning do worse (as of October 2026, Chess LLM Benchmark). Why this happens is explained in Why does ChatGPT play illegal moves?
• Facts about your own game. It can misremember which piece stood where, even in a game you just pasted.

How to verify: paste the position into the Lichess analysis board, whose computer analysis is powered by Stockfish (Lichess FAQ), or use Chess Coach in Claude or ChatGPT, described at the end of this guide.

Prompt 1: the ChatGPT chess coach prompt (set it up once)

This prompt turns ChatGPT into a chess coach with clear limits. Paste it at the start of a new chat, or save it wherever your assistant keeps standing instructions.

You are my chess coach. I'm rated about [1100] on [Chess.com] and I mostly play [rapid].
My goal for the next [8] weeks: [stop losing pieces in the middlegame].

Rules for every answer:
- Never invent evaluations, engine numbers or long variations. If a claim depends on
  calculation, say "check this with an engine".
- Only discuss positions I give you as a FEN or a move list. If anything is unclear,
  ask me instead of guessing.
- Explain ideas in plain words first. Use chess notation only in brackets.
- Keep answers short: three key points, then one question that makes me think.

Reply "Ready" and ask me what I want to work on today.

Why it works: the rules turn the model's biggest weakness, confident guessing, into an instruction to stop and ask. The short format keeps you thinking instead of scrolling.

Prompt 2: a study plan you can actually follow

Using ChatGPT as a chess coach works best for planning, because a plan needs no calculation at all.

Build me a 4-week chess study plan.
- Time: [30] minutes a day, [5] days a week.
- My rating: [1100]. My main problems, from my game reviews: [I hang pieces;
  I don't know what to do after the opening].
- Split each session into tactics puzzles, reviewing one of my own games with an engine,
  and one short theory topic.
- Name exact activities ("15 minutes of puzzles on Lichess", "replay my last loss
  with the engine"), not vague advice.
- Don't recommend books or courses unless I ask.
End with a one-line checklist for each day.

The last rule is there because a model can recommend titles or chapters that don't exist. If you want reading suggestions, ask for them separately and check each one.

Prompt 3: explain an engine line you paste

This is the most useful prompt in the guide. The engine supplies the facts; the model only translates them into words. Here is a real example from a well-known opening, the Two Knights Defense (1.e4 e5 2.Nf3 Nc6 3.Bc4 Nf6 4.Ng5 d5 5.exd5).

r . b q k b . r
p p p . . p p p
. . n . . n . .
. . . P p . N .
. . B . . . . .
. . . . . . . .
P P P P . P P P
R N B Q K . . R

Taking back the pawn with the knight feels natural. Stockfish disagrees. After that capture, White's knight can take the pawn on f7, next to Black's king, and attack the queen and the rook at the same time. That's a fork: one piece attacking two things at once. The engine rates the position +1.26 for White. Moving the other knight to the edge instead attacks White's bishop, which has no defender, and keeps the game about equal (+0.24). Now give the model those facts:

Here is a position from my game and what the engine says. Explain it to me like a coach.
Do not change or extend the engine's lines, and do not add your own evaluations.

Position (FEN): r1bqkb1r/ppp2ppp/2n2n2/3Pp1N1/2B5/8/PPPP1PPP/RNBQK2R b KQkq - 0 5
I played: 5...Nxd5
Engine reply to my move: 6. Nxf7 Kxf7, +1.26 for White
Engine's best move: 5...Na5 6. Bb5+ Bd7 7. Qe2 Be7 8. Nf3, +0.24 (about equal)

1. In plain words, what did my move allow, and what does the engine's move do instead?
2. What general idea should I remember for similar positions?
3. Give me one question to ask myself before I make this kind of move.

The engine supplies the moves and numbers, so the assistant has concrete evidence to explain. A prompt still cannot guarantee an accurate explanation: check that its description matches the board and the supplied line.

Prompt 4: study an opening through ideas, not long lines

Long memorized lines are where models slip most. Ask for plans and typical mistakes instead, and keep the move sequences short.

I play [the London System] as [White]. I'm rated [1100].
Teach me this opening through ideas, not memorized lines:
1. The 3 main plans for my side and the usual pawn structure.
2. Where my pieces usually go, and why.
3. The 2 most common mistakes players at my level make in it.
4. One typical trap for each side, described in words.
Keep every move sequence to 6 moves or fewer, and mark anything I should double-check
in an opening explorer or with an engine.

Check every line it gives you before you play it in a rated game. For your real results, Chess Coach's opening report (included in the Free plan) shows your repertoire with your scores and your weakest lines, based on your actual games.

Prompt 5: post-game reflection questions, before the engine

Writing down what you were thinking before you look at the engine lets you compare your reasoning with the truth, instead of just reading a list of moves. This prompt needs no chess knowledge from the model at all.

I just finished a game. Don't analyze it. Interview me like a coach, with 6 short
questions, one at a time:
- what I planned after the opening,
- the moment I felt the game turn,
- a move I spent a long time on, and why,
- how I used my clock,
- what my opponent was threatening that I missed,
- one thing I'd do differently.
After my last answer, summarize my thinking in 3 bullet points and tell me which moment
to check first with the engine.

Then open the engine and compare. The gap between what you thought and what was true is the lesson.

After the engine review: four follow-up questions

Once your game is finished and you have checked a key position with the engine, stay with that position. Ask these questions one at a time, waiting for the explanation before continuing. This extends prompts 3 and 5 rather than starting a new analysis from memory.

1. Understand the mistake: “What did my move let my opponent do?” In the Two Knights position above, the capture on d5 allows the knight fork on f7. Ask the assistant to point out the two attacked pieces on the board.
2. Understand the alternative: “What does the engine's move achieve instead?” Here, the knight move to a5 attacks the undefended bishop. Compare the purpose of the two moves, not just their evaluations.
3. Check the opponent's reply: “What reply does the engine show, and what must I notice?” The supplied line starts with the bishop moving to b5 with check. Follow that checked line; don't ask the model to invent the next ten moves or assume the opponent must play it.
4. Keep one usable lesson: “What should I check before a similar move in my next game?” For this example: before automatically taking a pawn back, look for the opponent's checks, captures and threats, including forks.

This is a worked example using the position above, not a transcript of a connector session. For your own review, supply the actual position and engine line. If an explanation mentions a move or threat outside that evidence, check it on the analysis board before keeping it in your notes.

What to ask an AI chess coach after a game: a real example

The Two Knights position is a textbook case. Here is the same routine on a real finished game from the founder's Chess.com account: nacsol3 (1223) vs aminbd1 (1249), May 5, 2026, a 10-minute London System. White was slightly better until move 18. The board shows that position, and every chess fact below comes from Stockfish.

r . . . . r k .
p . . b n p p .
. . . . p . . p
. . . p P . . .
. . p P q . . .
. . P . . N . .
P . Q . B P P P
R . . . . R K .

In the game, White moved the knight to attack Black's queen. It looked like pressure, but the two queens were already facing each other on the same diagonal, and it was Black's turn. Black's queen took White's queen, and nothing could take it back. Stockfish rates the position after the knight move as completely winning for Black. The engine's choice was to trade queens first. Here are the follow-up questions worth asking about that position, one at a time:

1. "Why was the queen trade better than my knight move?" Because trading removed the threat. Taking first meant White gave up a queen and got one back; the knight move left White's queen hanging. Ask the assistant to show both moves on the board, with the red and green arrows.
2. "What was the plan after the trade?" Ask it to follow the engine's line, not to invent one. Here, Black takes back with a pawn, which attacks White's knight. The knight moves and attacks Black's undefended pawn on the queen's side, and in the engine's line White's bishop wins that pawn. That is a plan you can describe in words: trade, then go after the weak pawn.
3. "What did my move leave undefended?" Besides the queen, the knight move stopped defending White's bishop. Ask which pieces each move protects and which it abandons. That is often the real cause of a blunder.
4. "What pattern should I practise?" One answer here: when the queens face each other, the side to move decides the trade. Before attacking the enemy queen with another piece, check whether it can simply take yours.
5. "Can I practise this position?" In Chess Coach, the board's Practice mistakes mode takes you back to positions where you went wrong, so you can find the better move without seeing it first. On Plus and Pro, drills collect such positions across your games. For other ways to turn mistakes into exercises, see chess puzzles from your own games.
6. "What one question will I ask myself next game?" Keep it short enough to remember during a rapid game: "Can their queen take mine right now?"

The answers in this list summarize the engine's findings in our own words. They are not a transcript of an assistant's reply. If an assistant's explanation of your game adds a move or threat that isn't in the engine line, ask it to check that position before you keep it. This game appears with the rest of that week in the founder's chess diary, and the hanging pieces guide covers the same blind spot.

Prompt 6: ChatGPT chess training with puzzles from your own games

Puzzles from your own mistakes are worth extra, because they show the patterns you actually miss. The trick is that the positions and answers must come from an engine review, not from the model. Collect three to five positions where you went wrong, with the engine's best move, and paste them in.

Quiz me on positions from my own games. For each one I give you the FEN, the side to
move and the engine's best move.
- Show me only the side to move and a one-line hint about the theme (for example
  "a piece has no defender").
- Wait for my answer. Don't reveal the engine's move until I answer or say "pass".
- If my answer is different from the engine's move, don't judge it yourself: tell me to
  check it on an analysis board.
- At the end, list the themes I missed.

1. FEN: [paste] | Best: [engine move]
2. FEN: [paste] | Best: [engine move]
3. FEN: [paste] | Best: [engine move]

The rule about different answers matters: sometimes your move is also good, and only an engine can tell you so.

Prompt 7: learn one concept, then check the examples

Concepts are a strength of language models. Examples are the weak point, so the prompt makes you verify them.

Teach me one chess idea: [the pin].
- Define it in two sentences.
- Explain how to spot it in my own games: what to look for on the board.
- Give me 2 practice positions as FENs, with the side to move.
- Then stop and remind me to paste each FEN into an analysis board to confirm the
  position is legal and the idea really works.

Positions made up by a model can be broken, for example with a missing king or a pawn on the first row. If a FEN won't load on the analysis board, ask for a new one or skip it.

Prompt 8: find patterns across your recent games

After a few engine-checked reviews, ask the model to organize your notes. It's good at summarizing text you give it, as long as it doesn't add chess facts of its own.

Here are my notes from my last [10] game reviews, all checked with an engine. Each line
has the result, the opening, the move where the evaluation dropped most, and my note.
[paste your notes]

1. Group my mistakes into at most 3 recurring patterns.
2. For each pattern, give me one habit to practice: a question to ask myself before moving.
3. Tell me which pattern to work on first, and why.
Don't add any chess facts that aren't in my notes.

How to check what the AI tells you

Every prompt above ends with the same step: put the position into an engine. There are two easy ways.

A free analysis board. Paste a FEN or PGN into the Lichess analysis board and turn on the engine. It's quick for a single position, but you copy and paste between tabs.

Chess Coach in Claude or ChatGPT. Chess Coach connects either assistant to Stockfish on our server. It retrieves your finished Chess.com or Lichess games, analyzes the key positions and lets you explore them on an interactive board. You can try your own moves and ask the engine to check the resulting position. The assistant explains the engine's findings; its wording still needs checking against the position and line. The Free plan includes one in-depth review and 30 position checks a day (docs). Practice mode revisits positions where you went wrong, which is prompt 6 without the copying.

Chess Coach is available in Claude’s directory and OpenAI’s plugin directory. Open its listing, install it and sign in with your Chess Coach account. Follow the setup instructions, and see the AI chess coach page for the review workflow.

Common mistakes when using ChatGPT as a chess coach

• Asking "what's the evaluation?" You will get an answer, and it will be a guess. Ask the engine instead.
• Pasting a whole PGN and asking where you went wrong, with no engine data. That's the setup that produced wrong suggestions in the Chess.com test above.
• Letting the model extend an engine line. The first moves are checked; the ones it adds are not. Prompt 3 forbids this on purpose.
• Accepting model-made puzzles without checking them. A wrong solution teaches a wrong pattern.
• Asking for help during a game. Chess.com and Lichess forbid outside assistance in games in progress (Chess.com Fair Play Policy, Lichess Terms of Service). Use these prompts only after the game is over.
• Chatting instead of training. A good prompt ends with something for you to do: a question, a quiz or a position to check.