Free chess game review: every mistake, explained

Review one game, or hand over the whole account

1. Review a single game

   Username, PGN, or game link. Full engine pass, no sign-up, nothing stored. Good for the game that just stung.

2. Connect your accounts

   Up to five Chess.com or Lichess handles by username. Public games only, no password, no OAuth. New games pull on open, roughly every six hours.

3. Stop asking for analysis

   Every game in your library is analyzed once, in the background, newest first. There is no separate training set to curate and no “analyze” button to remember.

What the engine pass produces

• Every move classified: Brilliant, Great, Best, Good, Inaccuracy, Mistake, Blunder, each with the centipawn loss and the move the engine preferred. Brilliant means a sacrifice that actually works; Great means it was the only move, confirmed by a deeper multi-line probe.
• Three accuracy numbers, all labelled: the headline is our centipawn-loss model, the same figure the library rows and the report card show. Below it sit a Lichess-style win-probability estimate and, on Chess.com games, their own score or a calibrated estimate of it. How accuracy works ›
• A one-sentence turning point: the move that actually turned the game, not the worst move on the sheet: it has to leave a tenable position no longer winning. A game lost on the clock is described as a game lost on the clock.
• Why the better move was better: a deterministic explanation per mistake, generated from the engine's own output and the detected motifs, never from a language model guessing at a position. On your own stored games, AI-written commentary is added on top and labelled as such.
• Master games from your positions: annotated games from the corpus that reached your structures, reachable from any key moment in the review.

The part that makes it a coach: mastery

Most sites hand you a percentage per game and let you draw your own conclusions. Elo Hell keeps a running statistical read on every skill you touch, updated by every game you play, so “what should I work on” has an answer with evidence behind it.

• Every skill carries a posterior: tactical motifs (forks, pins, skewers, discovered attacks, hanging pieces, back-rank and smothered mates), opening families split by colour, and endgame classes each hold a Beta distribution over how reliably you handle them, not a running average.
• Five plain tiers: Solid lately, Holding, Slipping, Leaking, Estimating. “Estimating” is the honest one: the model says so when it hasn't seen enough of a skill to have an opinion.
• Evidence is weighted by what it's worth: a real game counts double, a bullet game a quarter, a drill or an endgame play less than a game. Everything you do is evidence, but not equal evidence.
• It tracks form, not your lifetime record: evidence saturates at about 40 weighted observations (roughly 20 real games) per skill, and decays with a half-life of 360 days for tactics, 180 for openings and endgames. Tiers demote as readily as they promote, including when your rating climbs and the same skill stops holding up.
• One ranked answer: the Focus card names the single skill worth the most to fix and shows its case: how often it came up in your real games, what it typically costs you in pawns, and where it ranks among your tracked skills.

The full write-up of the model ›

Training that comes out of the model

• Puzzles from your own blunders: positions you actually got wrong, verified by the engine to have a single forcing solution that leaves you at least even, then scheduled by spaced repetition. Each day's set is drawn per skill from the mastery posteriors, so it follows what's currently slipping.
• Openings as a repertoire, not a quiz: build lines, review them on an Anki-style schedule, and see adherence and accuracy per family and colour from your real games, including the families you keep playing but have never prepared.
• Endgames played out: hold the verdict against the engine rather than find one move. Tablebase-exact to seven pieces with perfect resistance, 14 tablebase-certified theory drills, your own lost or squandered endgames extracted automatically, plus custom positions.
• A daily session that ends: own-blunder tactics, one shared Master of the Day, and master games similar to your own play. Three short sections, each resumable, any one of which keeps the streak.
• Statistics for the whole library: records, rating journey, rhythm and tilt, clock habits, where wins slip away, accuracy by move number. Every figure gated behind a real sample size.

What a free account gets

• A rolling library of 500 games: analyzed, coached from, and searchable. At the cap the oldest games rotate out; games you save are kept.
• 50 games analyzed per day: a first connect fills in about two days, after which you only ever analyze the handful of games you actually played.
• Up to five connected handles: Chess.com and Lichess, in any combination.
• No review tokens: reviewing a game you opened is not rationed against a daily allowance. The 50/day figure above governs background analysis of your imported library; guests and members alike are rate-limited only so the engine stays responsive.
• All of the training: puzzles, openings, endgames, mastery, and statistics are not paywalled features; the paid tier raises volume, it does not unlock the coach.