Mindset 不遷怒
The part that decides more outcomes than any opening, and the part with the worst evidence base on this site. Most chess psychology is one person's anecdote presented as a law. This page gives you concrete rules anyway — you still have to decide something on Tuesday night — but it says every time the ground is soft.
Asked which disciple loved learning, Confucius named one and defined him by two behaviours, neither of which is knowledge. 不遷怒 is not carrying anger into the thing that follows — the game after the loss. 不貳過 is not making the same mistake twice, which is the purpose of the log.
Tilt
The usable definition comes from poker research: a strong negative emotional state elicited by elements of the game — bad beats, a prolonged losing streak — that is characterised by losing control (Palomäki, Laakasuo & Salmela, Journal of Gambling Studies 2014, n = 417). Note two things. That is the definition the researchers handed 417 poker players before surveying them, not something they measured. And note what it does not say: it says nothing about decision quality degrading. An earlier version of this chapter added that clause — which is precisely the clause the rest of this section depends on, so it is now argued for below rather than smuggled into a citation. The mechanisms usually invoked — loss aversion, ego threat, and loss-chasing — are all borrowed. None has been demonstrated in chess.
What the chess data says, which is not what you have been told
Gee, Seese, Curley & Ward (2025, preprint) ran a hierarchical Bayesian analysis on Lichess bullet and blitz, asking whether recent results predict the next game. The population-level effect was centred on zero. Across individuals the change in win probability from a recent loss to a recent win ran −0.10 to +0.12, and most 95% posterior intervals contained zero.
Read that sample honestly, because the headline number flatters it. "Over 590,000 bullet games in the 1700–1900 cohort" is 36 players averaging more than 16,000 games each. The design randomly sampled 40 players per rating band plus 25 GM accounts — very deep per player, very thin across people, and selected for enormous bullet volume. It is a strong test of whether those players tilt and a much weaker one about everybody. It is also not yet peer-reviewed.
Against that: Chowdhary, Iacopini & Battiston (2023, Scientific Reports), 123 million Lichess games from ~0.98 million players, found streaks are real — longer and more frequent than a reshuffle of each player's own results predicts, cold streaks longer than hot, weaker players streakier. They could not identify the cause: rating advantage over the opponent correlated only ρ ≈ 0.28, time gaps between games ρ ≈ 0.05.
So streaks exist beyond chance, but the folk claim — this game is worse because I lost the last one — is unsupported at population level. Both datasets are online fast chess: Gee is bullet and blitz among a few dozen very active players rated 1700 and up; Chowdhary is blitz only, but across all levels. Neither touches rapid or classical. Nobody has measured a 1100-rated adult playing four rapid games on a Wednesday.
The damaging decision is probably not the quality of your next move. It is the decision to play the next game at all — and then the next four — when you have stopped learning from them. Rules work on that.
Set these before the session, not during it
A heuristic worth logging, not a validated measure: your average time per move over the first ten moves. When it drops well below baseline you are playing from memory rather than from the board — what tilt looks like from the inside, long before it feels like anger.
The plateau and the flat year
The 25th percentile of 1200-rated players studying eight hours a week gained exactly zero across a year. One well-documented adult improver went 1474 → 2007 in chess.com blitz over 10.5 years, with roughly four and a half productive years against six of flat or negative growth. Blitz, on a page that elsewhere treats blitz as a poor improvement vehicle — the shape of the trajectory is the point here, not the format. Lichess's trajectory analysis found players between 1600 and 2000 typically need four years or more to gain 100 points. Expected annual gain from 1100–1699 is about +121, and an average of +121 with any real spread around it necessarily contains a great many people at zero. A flat year is not the tail of the distribution. It is a normal draw from it. The thing that has not happened is that you have not been found out.
What to actually do during one
- Change exactly one measured variable, and give it twelve weeks. Changing five teaches you nothing about any of them, and it is what almost everyone does.
- Switch the scoreboard to process metrics, which move when rating does not. Pick two or three: games analysed engine-off per week (Principle Ⅱ); consecutive games without losing material to a one-move tactic; proportion of losses where you can name the threat you missed; confidently-wrong answers found per week (Principle Ⅳ); percentage of clock left at move 20.
- Check whether you are flat or measuring the wrong thing. Someone stuck at 1250 all year who went from naming the missed threat in 5 losses out of 20 to naming it in 19 has moved a long way.
- Consider that it is the ceiling of your time budget rather than a plateau. At two hours a week the arithmetic on the plan page is what it is. That is a legitimate place to be.
The move that turns a flat year into a flat three
Broadening. A new opening, a new course, a second app, a strategy book you will not finish. It produces the sensation of effort and is the reliable way to make no progress on anything for a long time. 弗措也 — the stopping rule from the Five Gates — points the other way: stay on the one job your rung names until its exit criterion is met.
Rating obsession
Elo's model assumes performance in a single game is normally distributed with a standard deviation of 200 points. Take that at face value and the standard error of your average performance over n games is 200/√n:
| Games played | Standard error of your measured performance |
|---|---|
| 10 | ±63 points |
| 30 | ±37 points |
| 100 | ±20 points |
Arithmetic from a modelling assumption, not a measurement — real chess performance is known not to be exactly normal. The order of magnitude is the robust part: a ten-game session cannot distinguish a real 60-point improvement from nothing at all.
The meaning of your rating has also changed without you playing a move: 1000 rapid sat near the 39th percentile in 2016 and the 81st by 2024. A rating target is a target on a number that wanders ±60 for free, which is why the Ladder uses exit criteria instead — "ten consecutive games without losing a piece to a one-move tactic" cannot be handed to you by a lucky week. Practically: stop looking between games; record the number monthly, or keep a 30-day rolling average. Honesty note: no study has tested whether hiding your rating improves results. It is a plausible mechanism and a widely reported experience, and that is all it is.
Playing scared
The pathologies
- Trading to relieve anxiety. The tell: you cannot say what the exchange achieved.
- Avoiding complications. Self-sealing — you never enter sharp positions, so you never build the patterns, so they stay frightening permanently.
- Resigning too early. Rung 3 exists because players at this level lose won positions. Your opponent will too.
- Draws offered from fear. If you cannot say why it is drawn, it is a resignation with better public relations.
- The merely-safe move, chosen because nothing can immediately punish it.
The fixes
- Pre-commit in writing before the game: "in this game I enter the complications." A decision made calmly beforehand survives the moment better than resolve summoned at the board.
- Check afterwards whether the scary line was actually bad. Engine off first, then on (the protocol). Most improvers find the line they feared was fine, and that does more than encouragement.
- Never resign a position you could not win from the other side against a peer. Play lost endgames out — free practice against a live opponent.
- Say the reason for a trade out loud first. No reason, no trade.
- Judge a loss by whether you can name the error, not by who beat you.
Evidence status: thin. None of this has been measured in amateur play. The nearest thing to evidence that who you are facing changes your move quality is Backus, Cubel, Guid, Sánchez-Pagés & Mañas (2023, Quantitative Economics): 28,799 tournament games, 2,504 players, engine-assessed over moves 15–30. Women's mean error was about 11% larger against male opponents (95% CI 5.4–16.9%), strength controlled. Two qualifications belong with it. It is observational, not an experiment — opponent gender there is real rather than manipulated, so it cannot separate what you believe about your opponent from who your opponent is. And it is contested: Stafford (2018, Psychological Science) analysed 5.5 million tournament games and found women outperform their expected score against men, with no stereotype-threat effect at all. Stafford does not stand unopposed either: Smerdon, Hu, McLennan, von Hippel & Albrecht (2020) published a reanalysis in the same journal reporting typical stereotype-threat effects after all. The studies also measure different things — move quality versus results. But this literature is not settled, and the honest reading is that "playing the person" is a coach's observation with a plausible mechanism, not a demonstrated effect.
Losing to lower-rated players
An opponent 200 points below you is expected by the Elo formula to score about 24% — one game in four. At 100 points below, 36%. Those losses are the model working as designed, not evidence about you. What they carry is ego damage, and ego damage is what breaks the stop-loss rule twenty minutes later.
Time trouble is usually not a calculation problem
Most improver time trouble is decision avoidance. You have two candidate moves, you have calculated both, and you are now spending four minutes trying to make the risk go away — which is not something minutes can do.
That less time costs accuracy is well measured at the top: in Chabris & Hearst's study of 23 grandmasters, blunders per 1,000 moves ran 5.02 at classical and 6.85 at rapid, and were worse when they happened. Taking the board away entirely cost almost nothing by comparison — which points at the clock rather than the board as the thing doing the damage. Calculation gives the full figures and what they do and do not license. Two limits belong here: it compares time controls rather than time trouble within a game, and the players are grandmasters.
Four rules that reclaim clock
- The re-check rule. If you catch yourself calculating a line you have already calculated, play your first candidate. Re-checking almost never changes the move; it converts a decision problem into a clock problem, and then you lose to the clock instead.
- Criticality first. Before spending, ask whether the character of the position is about to change — material, king safety, a structure locking. If not, move. Calculation lists the five practical triggers.
- Budget and log it. At 15|10, reach move 20 with more than half your clock.
- Audit your two longest thinks after every game. Did either change the move you would have played in thirty seconds? Most players find almost none did.
No study has tested time-management training in amateurs. These are coaching heuristics with a plausible mechanism.
Over the board
A person opposite you, silence, no takeback, no next game in ten seconds. The fear is proportionate to the novelty rather than the difficulty, and it decays fast. Go to a club night before a tournament; play your first rated event at the shortest classical control available; enter a section where you are among the lower seeds, so a loss is the expected result (see Play OTB). FIDE uses a K-factor of 40 for new players until 30 games, then 20 — your first tournament is a measurement with almost no precision.
Adult-specific problems
The competence inversion
You are competent at your job and have probably not been visibly bad at anything in public for twenty years. Chess returns you to being a beginner in front of strangers, with a number attached that updates in real time. That is the source of most adult chess distress and it has very little to do with chess. Naming it stops you looking for a training fix to something that is not a training problem.
Comparing yourself to children
Boring reasons carry most of the weight: far more hours, coaching from the start, tournaments most weekends, a peer group, and no fear of looking stupid.
The measured age penalty is smaller than the folklore. In the amateur improvement dataset, after adjusting for starting rating and study hours, age costs about 3 points of annual gain per year of age — so a 42-year-old versus a 12-year-old is around 90 points a year. Set that against the +121 a year expected between 1100 and 1699 and it is not a small effect: on those numbers the adult gains at roughly a quarter of the child's rate. That is the honest reading, and it is worse than the reassuring version. What it does not say is that the adult gains nothing — a quarter of +121 is still a rung every few years. Caveat: descriptive survey regression, self-selected and self-reported, and the two figures come from the same weakly-sourced dataset. Not causal, and not precise enough to take the "quarter" literally.
The starting-age effect that is firmly established concerns titles. Gobet & Campitelli (2007), n = 104: the probability of reaching international level — grandmaster or international master — was .24 for those starting by age 12 against .018 for later starters. Two honest qualifications: it is GM or IM, not IM alone, and those are raw proportions, not figures adjusted for practice hours (the practice-controlled result in that paper is a separate set of partial correlations). That is a statement about the very top. Nothing in it says an adult cannot reach 1800–2000.
"I'm just not talented"
Partly true, and the site will not pretend otherwise. Practice explains about 26% of the variance in skill at games (Macnamara, Hambrick & Oswald 2014, 157 effect sizes), and documented hours to master level range 3,016 to 23,608 (Gobet & Campitelli). Someone will reach your rating on a fifth of your repetitions.
But that 26% describes variance between people who all practised; it does not say you cannot improve. The sentence is almost always deployed at 1100, where performance is fully explained by not having done the work yet — talent is a real constraint at the ceiling and a rhetorical one in the middle. The screening behind this site's timelines found over a thousand adults who gained 200+ points after 25.
A piece of folklore worth naming
Chess coaching leans heavily on growth mindset, and the measured effects are small. Sisk et al. (2018, Psychological Science): mindset and achievement, k = 273, N = 365,915, mean r = .10, with 58% of effect sizes not significantly different from zero; interventions, k = 43, N = 57,155, mean d = 0.08, with 86% not different from zero — d = 0.34 for low-SES students, 0.03 for middle- and upper-class ones. Yeager et al. (2019, Nature), pre-registered, 12,490 students randomised across 65 schools, found real but small gains in the pre-registered target group — the 6,320 lower-achieving students — of about a tenth of a grade point on a four-point GPA, and larger only where school peer norms supported the message. Believing ability is malleable is probably true and worth believing; adopting the belief is not an intervention that will move your rating. The half-step is.
Guilt about the time
A hobby does not owe anyone an output. But if the guilt is real, the fix is almost always a smaller and fixed commitment rather than a larger vague one. "Twenty-five minutes, after the kitchen is done, six days a week" is defensible to a household in a way that "I'll be on the computer for a bit" never is. Two hours a week is a real plan on the plan page with a slower timeline, not a failed version of eight.
Why people stop
There is no published study of why adult chess improvers quit. Nobody has measured it. What follows is inference from single cases and should be weighted accordingly.
The de la Maza case
Michael de la Maza was an adult improver stuck around 1400 who published "400 Points in 400 Days" and then Rapid Chess Improvement. His method — the Seven Circles — was one fixed set of tactical problems (the CT-ART 3.0 set) solved seven times over, each pass compressed into a fraction of the previous pass's time, and very little else. The software is named in Heisman's account; the "each circle in a fraction of the time" description is the one universally reported by people who followed the programme rather than a line quoted from the book. It worked: he won the Under-2000 section of the 2001 World Open 8–1 and reached 2041 USCF. Nobody has reproduced it since — Tactics covers the twenty years of failed replication. Then he quit competitive chess permanently. In Dan Heisman's account: "To get to the next level, I would have to really learn how to play chess — advanced strategy and the like — that's something that would take some real work, and I don't think I want to do it."
The reading offered here — and it is a reading, not something he said — is that the method and the ending were the same fact. It optimised one narrow variable to exhaustion, delivered the number, and left nothing behind. A method you have to hate works right up until it stops, and then it takes the hobby with it.
The contrast is the improver above, whose 533 points came in about four and a half productive years. The other six were the years in which he kept playing, which is the only reason the four could happen. In the screening of adults who gained 200+ points after 25, the factor named first was consistent play — "more important than studying, coaching, chess psychology, or anything else."
Habit mechanics
Principle Ⅶ gives the shape: twenty-five minutes six days a week beats two and a half hours on Sunday. What follows is how to make it survive a real year.
How long before it stops needing a decision
Lally et al. (2010, European Journal of Social Psychology) followed 96 volunteers for 84 days adopting one new daily behaviour. Median time to 95% of the automaticity plateau: 66 days, range 18 to 254.
Two caveats. Of the 82 who gave usable data, only 39 produced a good fit to the asymptotic curve the 66 days comes from, and the behaviours were eating, drinking and exercise, not cognitive training. Some habits never go automatic and have to stay scheduled. That is a fine outcome, not a character failure.
The missed day, and what streaks get wrong
From the same study, in the authors' own words: "missing one opportunity to perform the behaviour did not materially affect the habit formation process." The specific figure sometimes quoted alongside this — under half a point on the automaticity scale — comes from secondary summaries rather than the paper's own abstract; the direction is the solid part. Missing one day resets nothing. Nearly all the damage is in what you conclude from it — the "what-the-hell effect" (Cochran & Tesser 1996), where a small breach triggers abandonment out of all proportion. Studied mostly in dieting research; the transfer is the unproven part.
Which is the case against streak counters for habits. A streak runs on loss aversion, and that is why it is brittle: what gets you to sit down on day 40 makes day 41 feel pointless. Count sessions per month instead. "22 out of 30" survives a bad week. A streak does not. The one streak this site does keep — the clean-game counter in the log — is not a habit streak. "Ten in a row" is the thing being measured, the exit criterion for Rung 1, and there the reset is the measurement working rather than a motivational device.
The minimum viable session
| You have | Do this |
|---|---|
| 90 min | The full week's work as your plan specifies it. |
| 25 min | One game at 15|10, plus the engine-off notes on it. |
| 10 min | Analyse one already-played game engine-off — three critical moments, from memory. |
| 3 min | Re-solve three cards from the failed-puzzle deck, cold from the diagram. |
When the week collapses, cut in the order the plan page gives: passive input first — reading, videos, anything new — then the extra long game, then puzzle volume but never to zero, and analysis last, because analysis is the feedback loop and the rest is unguided without it. honest tension That last clause conflicts with the finding above that consistent play was the factor named first in the adult-improvement screening. Both are defensible and this site will not pretend the conflict is resolved. If you can protect only one, protect whichever you have historically been worse at protecting.
Restarting after weeks or months away
When stopping is the right answer
This needs saying plainly, because nothing else on this site says it. If chess is producing real and persistent distress — if you consistently finish sessions angry, if losses follow you into the rest of your day, if you are concealing how much you play, if you have stopped enjoying any part of it and continue because of what you have already put in — then stopping, or stopping for a season, is a legitimate and healthy choice and not a failure of character.
No study will tell you when. But notice what this site's arithmetic implies: improvement is slow, largely unrewarded in the short term, and frequently flat for a year at a time. That is a reasonable deal for a hobby and a very bad deal for a source of self-worth.
There is also a middle option people forget exists: stop improving and keep playing. Rung 6 makes the same point from the other end — nearly all the pleasure of chess is available at 1800 and does not increase at 2000. And if what you are carrying is bigger than a hobby-sized problem, it is not a chess problem, and chess advice is the wrong tool for it.
不遷怒 — do not carry the last game into the next one, and set the rule that ends the session before the session starts. 不貳過 — convert every loss into one written rule. Expect the flat year, change one variable at a time, measure the process rather than the number, and turn up in twenty-five minute pieces.