The AI assistant
An AI assistant for chess academies
This AI copilot for chess coaches drafts game explanations and report narratives from source records, and your staff review them. It asks first by default before changing a record.
AI chess coaching for academies, on this page, helps your staff prepare and review work. Students practise against the engine in the student portal instead; see student training.
Models, allowance and limits
Three things to know first. The sections below cover approvals, missing data and data handling.
Which models
ChessCore's AI providers are Google (Gemini), Anthropic (Claude) and OpenAI. The privacy policy lists them, and how student data is protected explains what is masked first.
The daily allowance
Each academy has 200 AI copilot turns a day, where a turn is one request to the copilot. The allowance is shared across the academy, resets at midnight UTC and is the same on every plan including Free. See the daily AI allowance on every plan.
At the cap, and other limits
When the 200 turns are used up, the copilot pauses and resumes the next day (UTC). A missing roster or record is covered under what happens when data is missing, and marking invoices paid under actions requiring approval.
What the engine, the AI, the schedule and your staff each do
ChessCore keeps four jobs apart: engine analysis, written drafts, practice schedules and record changes. You can see who supplies the facts and who approves the words.
Engine facts and database numbers
Stockfish and your records
Every chess evaluation comes from Stockfish. Attendance, ratings, fees and dates come from your academy records. The language model does not judge a move and does not supply a figure.
AI-written explanations
The copilot writes the words
It turns engine lines and record facts into plain sentences for a student or a parent. By default this is a draft: a person on your staff reads it, edits it and decides whether it goes out. The one exception is class summaries, which an academy can choose to auto-publish when the coach notes behind them pass a completeness check.
Adaptive practice
A schedule, not a chatbot
Puzzle sets are ranked against a student's weak themes, and the due-now queue follows a spaced-repetition schedule. That is ranking and scheduling code. It does not write text.
Operational actions
Changes to your records
The copilot can propose a change, such as a new invoice or a rescheduled class. By default it asks first. Marking an invoice paid, refunds and payroll runs need a second person. The approval section below spells out each case.
Every draft, and the person who approves it
Pick a workflow. Each one shows the AI draft, the human edit and the click that sends it. Synthetic examples from Knight's Gate: the academy and everyone named are invented; these are authored illustrations, not captured AI output.
Game review for Mateo, game 1
AI draftApprovedAI draft
“Review the position before choosing your next move.”
Hannah edits it
“On move 5 your knight was pinned to your queen, so moving it let White take your queen. Before you move a piece, check whether it is pinned.”
Game review
Game review: Hannah adds the position
Until Hannah approves it, the explanation does not go to Mateo or to his parent. Every chess evaluation in it comes from Stockfish.
Progress report narrative for Mateo
AI draftApprovedAI draft
“Mateo is making good progress in attendance and rating, with games recorded for the month.”
Hannah edits it
“Mateo attended 7 of 8 classes, and his Lichess rapid rating moved from 1120 to 1185. His recorded games give us material to review together.”
- Narrative reviewed
- Numbers verified
- Ready to send
Progress report
Progress report narrative: Hannah replaces vague praise. Approve and send stays locked until she ticks Narrative reviewed, Numbers verified and Ready to send. The automatic monthly email is a separate fixed template with no AI text.
Report drafts see a first name and last initial. No AI report draft is generated for a student without active data-collection consent.
Follow-up to Chidi, new enquiry
AI draftSent by LaylaAI draft
“Hello Chidi, thank you for asking about Beginners A for Isla. Let us know when you would like to visit.”
Layla edits it
“Hello Chidi, would Monday at 17:00 at the centre work for Isla’s Beginners A trial with coach Daniel?”
Enquiry follow-up
New enquiry follow-up: Layla adds the trial details
Follow-up drafts for new enquiries stay drafts. A person on your staff edits them and sends them.
Personal fee note to Omar
AI draftSent by LaylaAI draft
“Hello Omar, Sara’s fee is overdue. Please clear the balance promptly.”
Layla edits it
“Hello Omar, thank you for paying USD 60 towards Sara’s USD 120 monthly fee on the 9th. USD 60 remains; please let me know when you expect to pay it.”
Fee note
Personal AI fee note: Layla reviews and sends; automatic fee reminders are fixed templates with no AI text
Marking Sara A.'s remaining USD 60 as paid is a separate, high-stakes change. Layla can ask for it; a second person approves it.
Copilot answer, forwarded to Daniel
AI draftForwardedAI draft
“Beginners A has 16 students, Intermediate B has 14, and the Tournament squad has 12. Knight’s Gate has 42 active students overall.”
Hannah edits it
“Daniel, Beginners A has 16 students and meets at the centre on Mondays and Wednesdays at 17:00. You coach this batch.”
Owner answer
Owner’s copilot answer: Hannah tightens it before forwarding to Daniel
The answer is a lookup. It reads the batch records and changes nothing.
Proposed change to the Beginners A timetable
ProposedChangedCopilot proposal
“Reschedule the Monday 17:00 Beginners A class to Tuesday at 17:00.”
Layla checks it
“Reschedule the Monday 17:00 Beginners A class to Tuesday at 17:00 at the centre, with coach Daniel.”
Schedule change
Schedule change: Layla asks the copilot to move one Beginners A class. Rescheduling is an ordinary change, so it waits for one click from her.
If Layla switches on auto mode, or allows that specific action, it runs without the click. After the change, the copilot looks the records up again.
Drafts coaches review: a game review for Mateo
One illustrative workflow at Knight's Gate, a made-up example academy. Mateo and Hannah are invented. The made-up example game was measured locally; the feedback and approval are illustrative, not a captured model response or an actual coach approval. Download the sample games (PGN).
- Stockfish
In this made-up example academy, Hannah selects a game for analysis before asking for a review for Mateo. In game 1, 5...Nxe4 changes the evaluation from -0.24 to +8.96 pawns from White’s view, measured locally on 11 October 2026 with Stockfish 19 at depth 16.
- AI draft
For a selected, analysed game, the copilot can draft an explanation from the engine output. In the made-up game, the f6 knight is pinned to the d8 queen by the bishop on g5. After 5...Nxe4, White takes the queen with 6.Bxd8; 6...Nxd8 then captures that bishop. The wording below is authored illustrative feedback, not a captured model response.
- Coach
Hannah edits one sentence, shown below, and approves the review. Until she does, the explanation does not go to Mateo or to his parent.
- Record
The draft, the approval and the send are on the record. No invoice, attendance mark or other academy data changed along the way.
What the AI drafts
- Game reviews, built from what Stockfish found in a game.
- Progress report narratives, written around real attendance, game and rating data.
- Staff review AI chess academy email drafts and parent notes before sending. Automatic monthly report emails use a fixed template with no AI narrative.
- Follow-up drafts for new enquiries.
The engine side of the review is described on the game analysis for coaches page. Monthly reports carry game counts and ratings, not engine analysis; see recurring student progress reports.
Adaptive practice is a separate system. A coach can rank the puzzle library against a student's weak themes when assigning a set, and the student's due-now queue follows a spaced-repetition schedule. See student training and homework.
What the assistant can do, and what needs a click
The person asking sets the limit. The copilot can do only what that person is allowed to do. Every action it takes is classed as a lookup, a change or a high-stakes change, and some changes count as high-stakes depending on what they touch.
| Kind of action | Example | What happens |
|---|---|---|
| Lookup | "How much is outstanding in fees?" or "Show Sara A.'s attendance summary." | Runs straight away. Nothing in your records changes. |
| Change | Ask the copilot to reschedule a class, or to create an invoice. | Waits for one click from the person asking. If that staff member switches on auto mode, or allows that specific action, it runs without the click. |
| High-stakes change | Mark an invoice paid, issue a refund, prepare a payroll run or remove a staff member. Creating an invoice is an ordinary change, not a high-stakes one. | Never runs on its own. It goes to an approval queue, and an owner, admin or finance staff member other than the person who asked must approve it. |
| Retry a workflow checkpoint | An approved step fails before the action starts. | Correct the inputs or permissions and request a new human approval. If the outcome is unknown, retry and restart are blocked: review the affected records and approval audit first. |
| Cancel a workflow | Stop a running workflow you started. | Cancellation is recorded and ends that workflow. It does not undo actions already completed; reconcile an unknown outcome before cancelling. |
| Only one approver | You are the sole owner and the only staff member able to approve. | You still cannot approve your own high-stakes request. Another eligible owner, admin or finance staff member must approve it; sole ownership does not bypass the rule. |
Who can use it. By default, academy owners and coaches can use the assistant. An owner grants access before admins, operations or finance staff can use this AI assistant for chess administrators. A coach's assistant works within the coach's own permissions.
Approving is a separate permission. Owners, admins and finance staff can approve high-stakes requests, and nobody can approve their own.
Questions an owner can ask
Ask in plain words about students, attendance, leads, invoices and outstanding fees. The copilot looks the answer up in your academy's records and reports what it finds. These examples use Knight's Gate, a made-up academy.
Answers come from your academy's current records at the time you ask, not from a nightly snapshot. After the assistant makes a change, it looks the records up again.
Made-up example academy, Knight's Gate: “For Sara A., inv-001 bills USD 120, records USD 60 paid on 9 September 2026, and leaves USD 60 outstanding. Mateo S.'s inv-002 bills USD 120 and records USD 120 paid, leaving USD 0. Together these two invoices bill USD 240, with USD 180 paid and USD 60 outstanding.” This is a worked answer from made-up invoices, due 5 September 2026 (see how the examples are made), not a captured copilot result or an academy-wide total. It excludes the other invoices, overpayment, refund and unmatched payment in the test set, and excludes ChessCore subscription fees.
“Which invoices are open for Sara A.?”
Answered from: Fee records
The copilot lists invoices for a student. It reports what is on the record and changes nothing.
“Who missed a class in Intermediate B this month?”
Answered from: Attendance marks
It reads the marks your coaches entered, filtered by batch and date.
“How many enquiries are in each stage?”
Answered from: Enquiry counts by stage
A count by stage, with no names needed in the answer.
“What is Mateo's current rating?”
Answered from: Rating records
It reads the rating ChessCore holds. For why a game was lost, the answer comes from Stockfish, not from the model.
“How many active students and coaches do we have?”
Answered from: Academy snapshot
The academy snapshot counts students and coaches as they stand when you ask.
“What was our attendance rate over the last 30 days?”
Answered from: Attendance marks
A 30-day attendance rate for the whole academy, from the marks on record.
“How much is outstanding in fees?”
Answered from: Fee records
The total still owed on your invoices, alongside 30-day billing totals.
“How many invoices are in each status?”
Answered from: Fee records
A count of invoices by status, without opening each one.
What the AI drafts for coaches
Five kinds of draft save your coaches typing. Each one says who checks it before it reaches a family or a batch.
Class summaries
Once a class is recorded, ChessCore drafts a lesson summary from the transcript, including any homework set in class. By default the coach reviews it, edits it and publishes it. Your academy can choose to auto-publish summaries whose coach notes pass a completeness check.
Recaps for each parent
A parent's class recap is built from the published class summary for their own child, with classmates' names removed. Coaches can copy it or open it in WhatsApp to send to a parent who has opted in to recaps. If your academy switches it on, recaps can go to opted-in parents automatically on WhatsApp or by email when a summary is published.
Lesson-plan drafts
ChessCore drafts a lesson plan for a batch from the curriculum topics its students score weakest on. The plan is built from curriculum mastery scores, not from students' games, and it is put together by fixed rules rather than a language model. A coach reviews each draft; approving it assigns homework and a matching puzzle set to the batch.
Studies and syllabuses
Coaches can ask the AI to draft a study or a syllabus. Nothing is saved until the coach confirms it.
Homework from a class summary
Homework the AI picks out of a class summary can be assigned to the batch with one click by the coach. It is never assigned automatically.
Actions requiring approval
Three rules cover what leaves the building and what moves money. Examples use Knight's Gate, a made-up academy.
What the AI never does
- Judge a chess move or supply a figureStockfish and your academy records do that.
- Approve on anyone's behalfApproval decisions are always made by a person.
- Send a review, report narrative or parent message it wroteThose stay drafts until a person on your staff sends them.
- Assign homework on its ownA coach assigns it with one click.
- Fill a gap in the record with a guessAn unmarked class is not turned into an absence.
- Show output that does not fit its fixed shapeThe draft is discarded, not shown.
The approval queue
- Rule 1
AI drafts wait for a person
Game reviews, report narratives and parent messages that the AI writes stay drafts. A person on your staff edits them and sends them. Class summaries follow the same rule by default, unless your academy chooses to auto-publish the ones whose coach notes pass a completeness check.
Example: Sara A.'s personal fee note waits in the queue for Layla to edit and send.
- Rule 2
Marking paid, refunds and payroll need a second person
Marking an invoice paid, a refund and a payroll run requested through the copilot become pending approvals. Creating an invoice is an ordinary change and needs one click. Owners, admins and finance staff can approve, and nobody can approve their own request. ChessCore prepares payroll runs and pays coaches out once a run is approved.
Example: Hannah, the owner, has given Layla, the admin, use of the copilot. Layla asks it to mark Sara A.'s remaining USD 60 as paid. Hannah approves it. Layla cannot.
- Rule 3
Routine reminders are automatic templates
Class reminders 24 hours and 2 hours before a class, fee reminders on a schedule the academy turns on, and monthly progress reports when the academy switches them on go out as fixed templates. They carry no AI text.
Example: Sara A.'s fee reminder on the 6th was a template the academy had switched on.
Other changes the copilot proposes, such as creating an invoice or rescheduling a class, ask first by default and run after one click. A staff member can switch on auto mode or allow a specific tool, and those changes then run without a click. Approval decisions themselves never run automatically.
Data handling is covered on the security page. The binding notice is the privacy policy. To see how approval queues work in coaching software or how coaches use AI before and after class, read the guides.
How AI output is checked
What you can conclude from this page. Every worked example here uses Knight's Gate, a made-up academy. They are synthetic examples, not observed production accuracy, an error rate or a time saving. ChessCore has not measured an accuracy percentage or a time saving for its product workflows, so none is claimed.
What we test. We defined a synthetic evaluation on 11 October 2026. It covers game reviews, a PGN comment that tries to give the AI an instruction, reports with missing records, a fee note, invoice totals and a schedule change. For each run it records whether the task completed, factual errors, unsupported claims, sentences a person edited and minutes to an approvable result.
What must fail safely. The failure cases include obeying an invented 2400-rating instruction hidden in a game comment, guessing the two classes Leo has no attendance for, inventing an engine evaluation, or mixing tuition invoices with ChessCore subscription fees. These are test cases, not observed failure rates. We will publish results only with the product version, model, reviewer, original draft and approved text recorded for each run.
Where the chess numbers come from. In the made-up game for Mateo, 5...Nxe4 was measured locally on 11 October 2026 with Stockfish 19 at depth 16, one thread, a 16 MiB hash and one candidate line, with the hash cleared for each position. The evaluation moves from -0.24 to +8.96 pawns from White's point of view. That is a local measurement on a made-up game, not production ChessCore output or an AI-quality score. Download the made-up sample games (PGN). For the product workflow, see game analysis and engine settings.
Automated checks on every product change
Every change to ChessCore must pass automated checks that AI report text and assistant answers use only numbers found in the source data and leak no contact details.
Coaches can flag a bad answer
When an AI answer or draft is wrong, a coach can flag it for the academy to review. Our synthetic evaluation method and failure cases describe what we plan to measure.
What the assistant remembers
You decide whether the assistant remembers facts about each student, guardian or coach. Memory is switched on or off for each person separately, and your academy can erase what the assistant remembers about someone.
What happens when data is missing
A gap in the record stays a gap. ChessCore does not fill it with a guess.
A class has no attendance mark
The report counts only classes with a mark. An unmarked class is not turned into an absence, and a month with no marks shows a dash instead of 0%. Leo T.'s two unrecorded classes stay unrecorded.
No consent on file
ChessCore does not generate an AI report draft for a student who has no active data-collection consent, and a report email needs parent communication consent.
No guardian email linked
If a student has no guardian email linked, no report email is queued and the reason is reported.
The roster cannot be loaded
The copilot masks student names using your roster. If it cannot load the full roster, it does not answer.
The daily allowance is used up
Each academy has 200 copilot turns a day, the same on every plan. A turn is one request to the copilot. When all 200 are used, the copilot pauses until midnight UTC and resumes the next day.
The model returns something unexpected
Output is checked against a fixed shape. If it does not fit, the draft is discarded, not shown.
How the assistant handles your data
Personal details are masked first
Before a copilot question reaches the model, ChessCore masks email addresses, phone numbers, ID numbers, dates and roster names. Report drafts see a first name and last initial. Two exceptions exist: the text-rewrite helper sends selected text as it is, and class audio goes to a speech-to-text provider before any masking.
Lookups stay inside your academy
Records the copilot reads are scoped to your academy, and names and contact fields are replaced or removed before the model sees them.
Usage logs hold metadata
AI usage logs record which model ran and how many tokens it used, not the prompt text. Saved copilot conversations stay in your academy account so you can return to them.
A daily allowance caps usage
Each academy has 200 AI copilot turns a day, reset at midnight UTC. It is the same on every plan.
Approvals are on the record
The draft, the approval and the send are on the record.
Approvals never approve themselves
The copilot cannot approve on anyone's behalf. Approval decisions are always made by a person.
The demo form opens on this page. Tell us which workflow you want to see.
Your AI, your records
Connect your whole academy to Claude, ChatGPT, Gemini or Perplexity
ChessCore gives AI assistants an MCP server and a public REST API. Each connection uses a scoped API key or OAuth, and an assistant receives only the records your academy authorises it to read. Claude, ChatGPT, Gemini CLI and Perplexity each let you add a custom MCP server on the plans that offer it, as does any other assistant that supports MCP.
Granted by Hannah, owner
From the blog
AI in chess coaching, explained
Get started
Start building a smarter academy
We set up your academy with you: roster import, Lichess and Chess.com connections, batches, and fee plans.
- Tenant isolation
- Personal details masked for AI
- Guardian consent
- Full audit trail