Play SC:BW
Přihlásit se Registrovat se
Hrát Žebříček Návody Učit se AI agenti Historie zápasů Hráči Streaming
Přihlásit se Registrovat se

Theory

How to read a replay

One pass for flow, three good moves, three bad ones, one change.

Newcomer Learner Competitor Advanced

The short version

  • Watch it once for flow and do not pause. You are looking for the moment the game turned, not for mistakes.
  • Then find three things you did well. This is not encouragement. It is how you find out which of your habits are load bearing.
  • Then find three things you did badly, each with a timestamp and a specific alternative.
  • Then change exactly one thing, and play ten games before you change another.
  • Half the analysis happens before the game: what they have done before, what the map argues for, and what the counter chain reduces to.

The situation

You lost. You open the replay, you watch four minutes of it, you see the thing you already knew about, you close it.

Or the thorough version, which is worse: you watch the whole thing, you write down eleven mistakes, you resolve to fix all eleven, and next game you fix none of them because eleven resolutions is the same as zero.

A review is not an inventory. It is a search for the one change worth making this week.

The method

Pass one: flow, no pausing. Watch the whole game at normal speed with the observer view on both players. Do not stop. Do not take notes. You are answering exactly one question: at what moment did the game stop being even. Write down a single timestamp.

Almost always that moment is earlier than you remember, and almost always it is not the moment you lost something. It is a minute or two before, when a decision was made that made the loss available.

Pass two: three things you did well. Go back and find three. Be specific: "I expanded at seven twenty, before they could punish it" rather than "my macro was okay".

This pass sounds soft and is not. The things you do well are the things you will still be doing next month, so you need to know what they are before you start rearranging your habits around a mistake. Players routinely destroy a good habit while fixing a bad one, because they never identified the good one.

Pass three: three things you did badly. Each one needs three parts: a timestamp, what you did, and what you should have done instead. If you cannot write the third part, it is not yet a mistake you understand, it is a mistake you feel bad about. Leave it out and find another.

Then: one change. Pick the one that was closest to the turning point from pass one. Write it as a sentence you can say before the next game. Play ten games. Only then pick another.

By rank, because the useful questions change

Newcomer. Watch only your own base, at double speed, and answer two questions. When did I first stop making workers, and why. What was the one thing that decided the game. That is the whole review, and it is enough: one verdict and one reason is what the coaching model says a newcomer should leave with.

Learner. Add checkpoints. Pause at six, eight and twelve minutes and write down three numbers each time: workers, production buildings, army supply. Compare them against the same three numbers from a game you won. The gaps will be obvious and they will be the same gaps every time, which is the point.

Competitor. Review decisions, not execution. For each fight: who chose to engage, where, and what was the trade in minerals rather than in bodies. For each build decision: what did you know at the moment you made it, and what did the budget line in T4 say they could have had. The question is never "was that right", it is "was that right given what you knew".

Advanced. Numbers rather than prose. Spending over time, production utilisation, the action budget by category from T10, and the fight by fight value trade. Then the part that is still not automatable: compare your decision points against a professional replay in the same matchup on the same map and find the one where you would have chosen differently.

The working

Show the maths

Ten games is not a trend. This is the piece of arithmetic that saves the most wasted effort, so it is worth doing properly.

If your true win rate in a matchup is pp, then over nn games the share you actually win has a standard error of

se=p(1−p)n\mathrm{se} = \sqrt{\frac{p(1-p)}{n}}
(1)

At p=0.5p = 0.5, that is about sixteen points over ten games, eleven over twenty and five over a hundred. So a six and four is entirely consistent with being a forty percent player and entirely consistent with being a sixty percent player, and drawing any conclusion from it is reading noise.

Put a number on how much data you would need. To tell a fifty percent player from a sixty percent player with any confidence takes on the order of

n  ≈  (zα/2p0q0+zβp1q1)2(p1−p0)2  ≈  200n \;\approx\; \frac{\left(z_{\alpha/2}\sqrt{p_0 q_0} + z_{\beta}\sqrt{p_1 q_1}\right)^{2}}{(p_1 - p_0)^{2}} \;\approx\; 200
(2)
Two hundred games to detect a ten point difference. Your last ten games are an anecdote.

Two consequences. First, do not change your build because you lost three in a row: you have learned almost nothing. Second, and more usefully, this is exactly why the method above asks for one change and ten games rather than for a measurement. Ten games will not tell you whether the change worked. They will tell you whether you can actually do it, which is the question that is answerable at that sample size and is usually the one that matters.

The things that are measurable in a single replay are the deterministic ones: your float, your worker count at a fixed time, how many production cycles you missed, how many of your units were firing in a given fight. Those need no statistics at all, which is why the checkpoint numbers above are worth more to a learner than any win rate.

The half that happens before the game

The class's model for a final project was a replay analysis, and the striking thing about the one the lecturer demonstrated is how little of it was about the replay. Most of it was about what both players knew before the first worker moved.

Precedent. What has this player done before, and what do they have a reputation for. A player with a famous build has something to defend, which makes them more likely to bring it, not less. The lecturer's point was that a signature build persists in variations while a build that wows everyone once tends to fade, and that the persistence is the thing you should expect.

Ego, and the series. In a series, beating a player's signature build in game one is worth more than the game, because it demoralises. Losing to it hardens their resolve. So the value of a game is not one win: it is a win plus its effect on the next two.

The map as an argument. This is the part most players skip entirely. Where are the chokes, how wide is the entrance to a natural, do the bridges force movement into an S shape, is the short path a tight passage or an open field. Every one of those facts constrains what defence is possible, and therefore what the opponent has to prepare for, and therefore what you should expect them to have built.

The worked version in the lecture: on a map with a tight short path and wide entrances to the bases, a Protoss needs cannons, and then cannon placement becomes a real dilemma. Forward with a gateway, they stop zerglings cold and are badly placed against ranged hydralisks that can stand off and outrange them. Placed back, the gateway and nexus soak damage first, which is good against zerglings. So the Zerg denies the scout, precisely so that the Protoss has to prepare for zerglings. The map argument and the scouting argument are the same argument.

Counter chains reduce to timing. The best structural observation in that analysis. If the Protoss build runs on corsairs, and corsairs lose to hydralisks, and hydralisks lose to dark templar, then the whole branching tree collapses into one question: when do the overlords arrive. Send the hydralisks early and the army is too thin to break the cannons. Send the overlords late and the corsairs kill them first.

When you can reduce a matchup to a single timing like that, you have understood it, and most of what looked like a decision tree turns out to be one dial.

What the lecture said

The two replays a week rule is set in lesson 1 at 35:00, and it is the most ordinary and most effective thing in the course: watch yourself, every week, all term.

The model analysis runs for the last twelve minutes of the lesson 11 lecture, from 1:07:00. The lecturer takes a professional Protoss against Zerg game and spends almost all of it on precedent, psychology, map geometry and the counter chain, reaching the game itself only at the end. That order is the lesson: most of what decides a game is visible before the first unit moves.

Lesson 1 also has the line the whole method rests on: take losing in stride, and spend a moment after every game naming the one thing that decided it. T14 is about that.

Still true today

Checkpoint

  1. You lost six of your last ten in a matchup. What does that tell you?

  2. What is the first pass of a replay review for?

  3. Why find three things you did well?

Go deeper

Sources for this page