How AI Progress Report Drafting Actually Works

Four inputs make an AI progress report worth reading: the student's synced games, their attendance record, their rating history, and the coach's own one-line notes from game reviews. Everything specific and trustworthy in the final report traces back to one of these. If a draft pipeline cannot show you its inputs, you are not looking at drafting, you are looking at improvisation.

• Synced games: full game records pulled from Lichess (first-class sync) or imported read-only from Chess.com, analyzed by Stockfish so the narrative can point at real moves [2].
• Attendance records: classes scheduled, attended, and missed this period, taken from the register rather than memory.
• Rating history: the rating trend over the reporting period, verifiable against the platform's public profile and rating system [3][5].
• Coach notes: the one or two lines a coach jots during a game review, like 'rushed in time trouble again' or 'first clean conversion of a winning endgame'. These give the draft its judgment.

Just as important is what must not go in. The model receives no free rein to invent numbers: attendance counts, fee amounts, and rating figures are injected from database records into locked fields, not generated as text. And personally identifiable information is minimized before any data reaches the model. The pipeline sends what the narrative needs (recent results, the attendance summary, the coach's notes) and withholds what it does not, such as contact details, payment instruments, or anything about other students. Our security page describes the data handling that sits underneath this; the principle is that a drafting model needs chess context, not a family's records.

A grounded draft starts from records and is constrained by them; a chatbot starts from your prompt and is constrained by nothing. Pasting 'write a progress report for a 10-year-old chess student' into a general chatbot produces fluent text with no relationship to the student. The difference is not writing quality, it is evidence.

The last row matters as much as the first. A grounded pipeline ends in review by design: every draft waits for a coach to edit or approve it, which is the control the NIST AI Risk Management Framework describes as human oversight of consequential automated output [1]. We wrote a whole piece on why that approval gate is non-negotiable, so here it gets one sentence: nothing AI-drafted reaches a parent until a named human approves it.

The coach still does the part that was always the point: judgment. In practice, coaches edit three things in an AI draft. First, tone: the model writes serviceable neutral prose, and a coach who knows the family adjusts it, warmer for an anxious parent, more direct for one who wants the unvarnished view. Second, emphasis: the draft may lead with the rating gain when the real story of the month was the student finally slowing down in winning positions, so the coach reorders what matters. Third, the next-step goal: the model can propose one from the data, but committing a student to 'two rook endgame sessions before the March open' is a coaching decision, and the coach owns it.

What the coach no longer does is assemble facts. Looking up attendance, checking the rating change, remembering which games were reviewed: the draft arrives with all of that in place. The coach's two minutes go into the sentences a parent will actually feel, not the bookkeeping. If you want the structure those sentences should follow, our progress report template breaks down the sections parents respond to.

Naive AI reports fail in three predictable ways: generic praise, hallucinated specifics, and the wrong emotional register. Each failure traces to the same root cause, a model writing without grounding in the student's actual records, and each has a specific structural fix. Knowing the failure modes is the fastest way to evaluate any vendor's drafting feature, including ours.

Generic praise is the most common failure and the most corrosive. An ungrounded model, asked to write about a young chess student, reaches for the safest possible prose: 'shows great enthusiasm', 'is making steady progress', 'should keep practicing tactics'. Every sentence is defensible and none is about this child. Parents read two of these reports, notice they could be swapped between siblings without anyone noticing, and stop reading. The fix is input, not prompting: when the draft is built from this student's synced games and this coach's review notes, the model has something specific to say, because specificity was handed to it.

Hallucinated specifics are rarer but far more expensive. This is the report that confidently praises a tournament result that never happened, cites a rating the student never held, or congratulates a family on perfect attendance during a month with two absences. One such error in a parent's hands undoes months of credibility, because it proves the academy did not read what it sent. The fix is the locked-number rule: attendance, fees, and ratings are injected from records the model cannot edit, and chess claims are tied to Stockfish analysis of real games [2], so the model is never the source of a checkable fact. Rating figures stay verifiable against the platform's own published history [3].

The wrong emotional register is the subtlest failure. A student has a rough month: dropped sixty rating points, missed three classes, lost a playoff in tears. An ungrounded model, biased toward positivity, drafts something cheerful, and a cheerful report after a hard month reads as either dishonest or inattentive. Grounding fixes the inputs (the draft knows the month was hard because the data says so), and review fixes the rest: the coach reads the draft with the child's actual face in mind and adjusts. This is also why the approval gate exists as a separate control; our piece on the approval queue covers what that screen looks like in practice.

The honest unit of savings is not minutes per report, it is whether monthly reports happen at all. Written from scratch, a thoughtful report takes a coach fifteen to twenty minutes of looking things up and finding words. A batch of fourteen students is an evening, and an evening that recurs every month is the kind of commitment that quietly dies by the third term. The most common progress report cadence at academies that write manually is 'we used to'.

With drafting, the work changes shape. The facts arrive pre-assembled and locked, so review is reading, not research: a coach reads each draft in a couple of minutes, edits the few that need a tone or emphasis change, and approves the batch. In the demo academy, Coach Priya's review of Batch B2's monthly reports is coffee-break work rather than an evening, and the reports actually go out, every month, with her name on each one. The compounding effect matters more than the per-report arithmetic: consistent reports beat occasional brilliant ones, and consistency is what drafting buys.

Parents notice specificity, not authorship. No parent has ever asked whether a sentence was first typed by a model or a coach; every parent notices whether the report is about their child. A grounded draft mentions the actual game where the student held a worse rook endgame for thirty moves, the actual attendance figure, the actual rating trend, and those checkable details are what make a family feel seen.

Delivery is part of the impression. Reports and recaps reach families where they already are, over WhatsApp through the business messaging tools academies use [4], after approval. A parent who can open the report, tap through to the cited rating history, and find it matches the platform's public record [3] learns to trust the next report too. Specificity plus verifiability is the whole trick; our guide on showing parents rating progress goes deeper on making the numbers legible to a non-chess family.