AI for chess coaches: lesson prep and review

AI helps a chess coach most as a drafting tool. Before class it turns an engine-checked game into a lesson draft and a few board questions; during class it stays out of the way, because the coach is the one reading the student; after class it drafts recaps and report text that wait for the coach to approve them. This guide follows one worked lesson from a stand-in game to the parent update, using an invented academy and a made-up intermediate student whose Lichess rapid rating moved from 1120 to 1185 over the term. The game is a well-known short trap, replayed through a chess library and Stockfish on 16 October 2026 so that you can load every position yourself. None of this is a customer story.

The order matters. Evidence comes first, because an AI that is asked to find the mistakes in a game will sometimes invent them. The draft comes second, because a first draft is cheap to read and cheap to throw away. The check comes third, because that is where a coach earns the lesson. The assignment comes last, because a lesson that ends without practice is a conversation, not a plan. The four headings below follow that order.

This page is about the working method, and it applies whichever tools you use. If you want to compare consumer AI tutors for a player, see the player-facing apps compared. The structure of a lesson itself, with timings for each part, belongs to the chess lesson plan template; this guide covers only what AI adds before and after that plan.

A lesson draft is only as good as what the AI is handed. For the worked example, the evidence is one game and a short record. The game is a short loss with Black, listed move by move below; in real life it would be a PGN downloaded from the student's account, which any Lichess account can do for free [3]. The record says he attended 7 of 8 classes, that his rating moved from 1120 to 1185 over the term, and that the batch has been working on checking an opponent's threats. The coach hands the AI that and nothing else.

The stand-in game is 1.e4 e5 2.Nf3 d6 3.Bc4 Bg4 4.Nc3 g6 5.Nxe5 Bxd1 6.Bxf7+ Ke7 7.Nd5#, a well-known trap shape in which White leaves the queen on d1 as bait. The position after 5.Nxe5, with Black to move, is rn1qkbnr/ppp2p1p/3p2p1/4N3/2B1P1b1/2N5/PPPP1PPP/R1BQK2R b KQkq - 0 5. Paste it into the free chess board editor to follow along; the last position is checkmate.

app.chesscore.io/students/lucas-m/games/review

Knight's GateStudentsLucas M.Game review

Lucas M. v Kofi M.Lichess rapid 15+10 · 14 OctMissed mate

8♚︎7♟︎♟︎♟︎♟︎6♟︎♜︎5432♟︎♟︎♟︎♟︎♟︎1abcd♜︎efg♚︎h

Stockfish, depth 22 · Move 17BestRd8Mate in 1Playedh3+0.0

AI explanationDraft

On move 17 the rook could go to d8 and it was mate: the king had no square to run to. h3 was played instead, and the chance was gone.

Coach Elena edits and approves before it can be sharedEditApprove

Parent portalAcademy rating 1342WWLDWRecent results and rating, no engine text

Stockfish does the finding. The engine marks the moments where the evaluation swung, and in this game it flags Black's fifth move as the turning point. On 16 October 2026 a local Stockfish 19 run at depth 18 rated Black's best reply to 5.Nxe5, which is 5...dxe5, at roughly 2.4 pawns behind. After 5...Bxd1 it found a forced mate in two: 6.Bxf7+ Ke7 7.Nd5#. The coach can check this in seconds by loading the position and playing the moves. Engine output is checkable in a limited sense: rerun the position with the same version and depth and you should see a similar verdict, though exact numbers vary with depth, version and machine, which is why the version and depth are written down [1]. Stockfish is free and open source, and its official blog announced version 19 on 5 September 2026 [5].

Before you ask for anything, write down three things yourself. First, the position at the turning point, so you know what White's knight on f3 was pinned to. Second, what the student was probably thinking, which only you can guess from knowing him. Third, what you want from the lesson: one idea, not five. In the example the coach writes: he took the queen without asking what the capture allowed. That sentence is the brief for everything that follows.

Keep the evidence small. A whole month of games invites the AI to generalise, and generalisations are where it goes wrong. One game, one turning point and one pattern is enough for one lesson. If the same pattern shows up in two or three other games, that is a finding for the progress report, and you can confirm it from the engine marks on those games rather than from the AI's say-so.

Also keep private details out of what you paste into any outside tool. A student's full name, a parent's number and the fee record are not needed to explain a pinned knight. Use the first name and last initial, as this page does.

• Record: Lichess rapid rating 1120 to 1185 over the term; attendance 7 of 8 classes.
• Game to review: the stand-in loss with Black, with move 5 flagged.
• Question for the board: after the queen is taken, what does 6.Bxf7+ threaten?
• Coach check: replay the position and correct any draft that says the knight on f3 was pinned to the king.
• Planned follow-up: the coach selects 12 puzzles of the same pattern due Friday; completion is not yet claimed.

Give the AI checked evidence and ask for a draft in a specific format. In the example the coach asks for a short plain-language explanation of the turning point for a student rated around 1150, a three-step way to check what a capture allows, and two or three questions to ask the student at the board. The coach asks for a draft, not a finished lesson, and says so. The briefing below is a copyable template for the request.

• Position reference: the FEN or the move number, with the student's name removed.
• Checked engine observations: engine name, version, depth, the best move and the evaluation you replayed yourself.
• Student goal: the one habit this lesson is for.
• Terms already understood: the words you can use without explaining them.
• Questions requested: how many, and at what difficulty.
• Prohibition: add no moves, evaluations or conclusions that are not in this briefing, and say 'not provided' if asked about anything else.

Here is the kind of first draft the AI returns, shortened to its structure. It says that the turning point came on move 5, when Black took the queen on d1. It explains that White's knight had been pinned against the king, and that a pin means a piece cannot move without exposing something behind it. It proposes a habit: before taking material, ask what the opponent's last move leaves undefended or allows. It suggests that the student replay the position and find a better reply. The draft reads smoothly and is mostly right.

This is the point of a draft. It gives the coach something to react to. A blank page asks the coach to think of everything, and a draft asks the coach to notice what is missing, which is a faster and more reliable task for most people. It also tends to include a good question or two that the coach would not have written at that moment.

Do not ask the AI to also schedule, message or grade. Keep the request to the lesson. If your platform drafts the review for you, the same principle holds: the draft is a starting point, and it waits for a person.

This step is the whole job. Read the draft slowly, one sentence at a time, and test each claim against the board. The method is simple: if the draft names a move, play it. If it names a square, look at the square. If it names a rule or a principle, ask whether it applies in this position or only in general. The longer comparison of engine output and AI explanation lives in our guide to engine analysis versus AI explanation.

In the example, the coach finds three problems in the first draft. First, the draft says White's knight was pinned to the king. A knight pinned to the king cannot legally move, and the game record shows White moved it with 5.Nxe5, so the pin was against the queen on d1. That is a relative pin: the piece can move, at a price. The distinction changes the lesson. The student did not take advantage of an illegal move. He took a queen that White had left as bait. The coach checks the position and rewrites the sentence to match what happened.

Second, the draft suggests a habit that is correct in general but too long for a student at this level to apply during a ten-minute game. The coach shortens it to one question: what does this capture allow. Third, the draft suggests a different reply for Black. The coach replays it and finds that the engine prefers 5...dxe5, which leaves Black clearly worse but not lost, so the suggested reply is not the best one. The coach deletes that sentence and keeps only what the engine line supports.

The final plan is shorter than the draft. It says: the turning point is move 5, the knight was pinned to the queen and not the king, the lesson is to ask what a capture allows before taking, and the practice is a set of 12 puzzles due Friday. The coach has changed what the knight was pinned to, the habit and the suggested alternative. The AI supplied the structure and two of the questions, and the coach supplied the truth.

Use the same routine every time. Replay every line on a board. Check every named square and every named piece. Look for any claim that sounds like a general principle and ask whether this position is an exception. Read the tone as the student would read it, because a correct sentence can still discourage a nine-year-old. And never approve a draft you have not read to the end. Reading it quickly does not count as checking it.

Knowing where the tool fails saves more time than any prompt. The failures below are the ones coaches should expect, and none of them is a reason to stop using a draft. They are reasons to read it carefully.

• Engine lines it misreads. The engine's line is reliable, but the language model sometimes describes it badly. It may say a move is bad because it loses a piece when the engine only prefers another move by a small margin, or it may call a forced sequence a mistake. Read the engine line yourself.
• Quiet positions. When no move wins or loses material, the engine's preference between two good moves is subtle, and an AI explanation of it tends to sound confident and vague. Strategic plans, pawn structures and prophylaxis are common places for a draft to say something plausible and empty.
• Endgames. Exact technique, such as opposition, triangulation and rook endgames, depends on precise move order. A draft can describe the idea correctly and give the wrong square.
• Positions it explains badly for the level. A draft may use terms the student does not know, or explain the right move with a long line the student cannot hold in their head. Match it to the student, not to the engine.
• Anything about the person. Why a student played quickly, whether they were tired, whether they are losing interest, and what their parent expects are not in the game record. Only the coach knows.

A useful habit is to keep a short list of the errors you find in drafts. After a few weeks you will see which kinds of sentence you correct most often, and you can tell the AI up front to avoid them or simply read those parts with extra care. The list is your own, and it will not match another coach's.

There is also a limit on what you should hand over. The AI should not choose which student gets which lesson, decide that a student is ready for a tournament, or word a message about a missed payment. Those are the coach's and the academy's calls. If you want to see how an academy keeps that boundary in software, read our guide to AI guardrails in coaching software.

A lesson ends with practice that the student can do alone. In the example the coach assigns 12 puzzles of the same pattern, due Friday. That one assignment matches the one weakness the game showed. The coach did not ask the AI to invent a study plan, because the evidence supported a single pattern, and a single pattern needs a single, short assignment.

Set the homework for one student, or for the whole batch if another student showed the same pattern, with a due date the student can see. Puzzle sets work well here because they are short and checkable, and free trainers already serve them: Lichess puzzles built from real games are unlimited on a free account [3]. This is where the lesson becomes something the student does between classes, so the assignment should name the pattern, the number of puzzles and the day it is due.

app.chesscore.io/students/lucas-m

StudentsLucas M.

PASearch⌘K

AR

Lucas M.

Batch B2 · Thursdays · joined Aug 2025

+118 since June

1342

Academy rating · Lichess games synced

Games · 142Reports · 14FeesAttendance

Won vs. 1410 · French Defense

Best game

Missed fork on move 23

Homework set

Full history kept · survives coach changesEvery change logged · audit trail

Write the assignment so you can tell next week whether it worked. Twelve puzzles by Friday is checkable. Work on tactics is not. Then, when the student returns, look at the result before you plan the next lesson. If the student solved the set but took a similar bait in his next game, the lesson needs a different angle, and the engine marks on that game will show you where.

Some assignments should not come from the AI at all. Choose the difficulty yourself. A puzzle set that is too easy teaches nothing, and one that is too hard teaches the student to avoid it. You know the student, and a draft does not.

Keep a record of what you set, what came back and what you changed next. That record is what lets you say something true to a parent. It is also the honest basis for any claim about progress, because it shows what the student actually did rather than what a draft said they would do.

Lesson-prep software should let you bring your own verified evidence, edit every draft, and decide what is assigned. It should never send a lesson or homework the coach has not approved, and it should be honest about which parts are AI and which are rules. Ask any vendor three questions: where do the evaluations come from, what happens to a draft nobody approves, and can the coach change the assignment before it reaches the student?

ChessCore, which we build, is one example of how that can work. On the Pro plan and up it syncs public games from Lichess and Chess.com read-only, shows an analysis status for each game (Pending, Analysing, Done or Error), and holds AI-drafted text for a person to edit and approve [2]. For a whole batch, its lesson planner drafts a plan from the batch's curriculum mastery scores. That planner is rule-based, not AI, and it works from those scores rather than the students' games, so a game like this one is still something the coach brings. The draft waits for review; approving it assigns homework and matching puzzles to the batch, and nothing is assigned while it is pending. To plan a lesson by hand instead, start from the chess lesson plan template.

The same lesson can feed the monthly parent update. In the example, the student's report for the month has four facts from records: attendance of 7 of 8 classes, a rating that moved from 1120 to 1185 over the term, one recurring pattern, which is taking material without checking what the capture allows, and one next step, the 12 puzzles due Friday. The AI can draft the narrative around those facts. The coach edits it, then approves it.

Numbers in a report should come from records, not from the draft. If the draft says attendance was 8 of 8 and the record says 7 of 8, the record wins. A report should also not claim improvement that the data does not show. A rating rise is a fact, but a claim that the student has fixed the habit is not, until his games show it.

Two kinds of message are different, so keep them apart. AI-written report text is a draft that waits for a person to approve it. Automatic monthly reports, when an academy switches them on, send a fixed template with no AI-written text, and class and fee reminders can also send automatically on fixed templates. If you want a layout for the update, the progress report template shows the sections a parent update usually needs. If you want the longer treatment of how AI-drafted progress reports are built and reviewed, that has its own guide.