Building EloCoach
A public engineering notebook. What I built, what broke, what it cost, and what I changed afterwards.
Experiments
I cut the coach’s prompt in half. The first word still took 20 seconds.
A shorter prompt was not shown to keep review quality, and the wait is the model thinking. Updated with a stricter study of where the coach goes wrong, and the board facts we shipped.
13 min read
LLMs
Ask the coach why. Ask the board what if.
Explaining one blunder, my coach suggested three bishop moves that cannot be played. Why a language model misreads the board, what the coach is reliable for, and when to play the line out yourself.
6 min read
Engineering
The opening book skips moves. I built an opening trainer on it anyway.
The open Lichess opening list names only some positions on each line, so a first tree called real Ruy Lopez theory a miss. What I found, how I fixed it, and what the book still cannot say.
6 min read
Engineering
What is fixing one chess habit worth? How EloCoach estimates it
A rating promise on a button has to survive the player doing the arithmetic. How the losses one habit decided become a cautious Elo estimate, and which part is still a guess.
6 min read
Experiments
17 accounts, 162 games: what actually makes an EloCoach review slow
A baseline run on grandmasters and streamers from Chess.com and Lichess: engine time follows game length, coach time follows its own answer, and eight at once hits a credit wall.
9 min read
Engineering
How EloCoach fetches your games from Chess.com and Lichess, and what a 422 means
Per-user OAuth on Lichess, careful serial requests on Chess.com, the limits of each platform, why a fetch can end in a 422, and a thank you to both.
8 min read
Experiments
I won this game. Stockfish says I was losing for twenty moves.
The same 61-move game through raw Gemini and through EloCoach, with every board claim replayed in chess.js and an honest scoreboard where neither coach comes out clean.
10 min read
Experiments
I gave Gemini and EloCoach the same four-move game. Both got things wrong.
A 582-rated blitz loss, a bare PGN versus the full EloCoach pipeline, every claim replayed in chess.js, and what the errors say about where verification has to live.
9 min read
Infrastructure
What it costs to run EloCoach, and why the free tier needs a gate
Real numbers from a beta paid out of my own pocket, the actual Gemini bill, what changes when users show up, and how we plan to price it fairly.
8 min read
Product
The ten seconds before the board: making a game review feel like something
A result screen with a mood-matched GIF and a cheeky coach line, what I decided against, how it works, and eight sample GIFs for wins, losses and draws.
6 min read
Infrastructure
The boring stack that runs EloCoach, and why each piece is there
Koyeb, Postgres, Inngest, Resend, OpenRouter, PostHog, LangWatch, Expo, Adapty, Giphy, Cloudflare and GitHub: what each does and one diagram of the whole thing.
9 min read
Engineering
I swapped in Stockfish’s own win model. Two days later I took it back out.
Why a model fitted to engine games grades a club player’s one mistake as 100 to 0, and the three separate problems that turned up once I looked.
7 min read
Engineering
Maia 3, your rating, and the day its favourite opening move was h2h3
Why Maia 3 beat per-band models for coaching, how we adjust its rating input by time control, and the bug that scrambled every "human move" until a sanity check caught it.
9 min read
LLMs
We tried other models. Gemini is still the one, for now.
Prompt tests, a model that burned its whole budget thinking, and why the coach of the future might run on Claude or GPT when the coach says so.
8 min read
Engineering
What counts as a “brilliant” chess move? My five attempts at an answer.
From “is the piece attacked?” to walking the engine’s line until the board is quiet, and each time a wrong label reached a real game.
7 min read
Engineering
My coach kept inventing chess moves
Illegal lines, empty boards, and the unglamorous machinery (chess.js, validators, a rule about who writes last) that keeps an LLM honest about the board.
8 min read
Why we built it
Chess coaching is priced for the few. Most of us were never in the room.
Four-figure programmes, lessons from 30 to 200 dollars an hour, and what that does to everyone who just wants to stop hanging pieces.
5 min read
Why we built it
I pasted my games into Gemini for a month. Then I built what I actually wanted.
The surprisingly good, deeply inconvenient reviews that started EloCoach, and why I wanted a coach that remembers me instead of repeating "watch out for blunders".
6 min read