Quick answer
Quick answer: Take the knight when it gives check or a fork the queen cannot, and take the rook when a queen would stalemate.
Every other time, promote to a queen without thinking about it.
Underpromotion is rare on purpose. It exists for two specific problems, and outside those two it throws away material for nothing.
So this isn't a habit to build. It's a check to run in the handful of positions where the automatic move loses.
The rule allows all four
Worth stating plainly, because beginners often think otherwise.
A promoting pawn can become a queen, rook, bishop or knight, and the choice is entirely yours.
You are not limited to pieces that have already been captured, so nine queens on the board is legal if you can get there.
The promotion rules confirm the new piece does not have to be a previously captured piece.
That last point matters at a club, where somebody occasionally objects because there's only one queen in the box.
Why the knight is the useful one
Because a queen can do everything a rook and a bishop can do, and nothing a knight can.
That's the whole logic of underpromotion in one sentence.
A rook promotion gives you less than a queen in every direction. A bishop promotion gives you even less.
A knight promotion gives you a jump the queen cannot make, and that jump is sometimes check, sometimes a fork, sometimes mate.
So when you see a promotion square, the only piece worth considering against the queen is the knight.
The knight fork pattern
The most common practical case, and it's a shape you can learn.
You promote on the eighth rank, and the new knight immediately checks the enemy king while also attacking a queen or a rook.
The opponent must answer the check, and you take the second piece next move.
That turns a promotion into a promotion plus a free queen, which is a much bigger swing than the difference between a knight and a queen.
Our tactics guide covers the fork shapes worth recognising, and this is the same geometry arriving from the pawn.
The defensive knight promotion
Less known, and it saves half points.
In a rook against pawn endgame, promoting to a knight is described as a standard defensive technique.
You're losing anyway with a queen, because the rook picks it off or mates. A knight can check, jump away and survive.
It's not a win. It's a way to keep playing when the automatic move ends the game immediately.
Our rook endgame guide covers the family of positions this comes from.
Promoting to a rook to avoid stalemate
The second reason, and the more elegant one.
Sometimes a new queen takes away the last square the enemy king has, and the game is drawn on the spot.
A rook covers less, leaves the king a move, and wins.
This is the classic underpromotion in study composition, and it shows up in real endgames often enough to be worth the check.
Our stalemate explainer covers how easily a winning position turns into half a point.
The Saavedra position
The most famous example in chess, and it's worth playing through once.
White has a king on b6 and a pawn on c7. Black has a king on a1 and a rook on d5. White to move.
Promote to a queen and Black plays Rc4 with check. Take the rook and it's stalemate, so the win evaporates.
Promote to a rook instead and White threatens Ra8 with mate. Black defends with Ra4, White plays Kb3, and Black loses the rook or gets mated on c1.
The full study was published in Glasgow in 1895, and the winning move came from a priest who was a weak amateur.
I read the study through move by move rather than trusting the summary, because the stalemate line is the only part that explains the choice.
The Lasker trap
An opening line where an underpromotion appears on move seven, which is absurdly early.
It arises in the Albin Countergambit, where Black's advanced pawn reaches the second rank and promotes to a knight with check.
A queen there simply loses. The knight wins material.
It's the standard answer to anybody who claims underpromotion never happens outside compositions.
You'll almost certainly never see it in your own games, and it's still worth knowing that a move seven underpromotion is a real line.
The bishop promotion
Essentially never, and it's worth saying why.
A bishop can do nothing a queen cannot, so promoting to one is strictly worse in every position except stalemate avoidance, and a rook handles that better.
It's described as rarer than knight or rook promotion, which is already rare.
Where it appears is in composed problems, particularly the ones showing all four promotions in one solution.
In a real game, if you're reaching for a bishop, check your reasoning twice.
The joke promotions
Worth naming so you don't copy them.
Many practical underpromotions have been called silly jokes, and that's a fair description.
A player who's completely winning promotes to a knight because it's funny, and the game was over either way.
That's fine between friends. It's a bad habit in a rated game, where a needless handicap occasionally turns a win into work.
Promote to a queen unless there's a reason. Being amused is not a reason.
The check to run
Three seconds, every time a pawn reaches the seventh.
First, does a new queen leave the enemy king with a legal move? If not, look for a rook promotion.
Second, does a knight on that square give check? If it does, look at what else the knight attacks.
If both answers are no, take the queen and move on.
That check costs you nothing across a hundred games and saves you the one game where the automatic move draws.
Promoting more than once
A situation people handle badly under time pressure.
Two queens are fine, and so are three. There's no rule limiting you to the pieces in the box.
Over the board, ask the arbiter for a second queen or use an upturned rook by agreement, and never just leave the pawn on the eighth rank.
A pawn sitting on the last rank is not a queen and creates a dispute you will lose.
I checked the wording on that specifically, because it's the kind of detail a first tournament punishes.
How to practise it
This is a pattern recognition skill, not a calculation skill.
Promotion puzzles come up regularly in tactics training, and they're distinctive once you've seen a few.
The tell is a pawn on the seventh with an enemy king that has almost no squares, or an enemy queen sitting a knight's jump from the promotion square.
See that shape twenty times and the check becomes automatic.
You can drill exactly this on the puzzles, where promotion positions appear often enough to build the reflex.
Online play makes it easy to miss
A practical warning for anybody who plays mostly on a screen.
Most sites auto promote to a queen, or make the queen the default in the picker, so your hand moves before your brain does.
In a bullet game that's correct. In a slower game it costs you the one position a year where the queen is wrong.
If your site has a setting for it, turn auto promotion off for longer time controls and leave it on for bullet.
That single setting is the whole practical value of this article for an online player.
It is a losing-position tool too
Worth flipping the perspective, because most people only think about winning.
When you're worse, a promotion that gives check buys you a move, and a move is sometimes a perpetual or a fortress.
A queen that gets captured immediately buys nothing at all.
So in a lost position the question changes from which piece is strongest to which piece survives the next move.
That reframing is where most defensive underpromotions come from, including the rook against pawn case above.
The short version
The queen is right almost every time, and you should promote to one without hesitating.
Take a knight when it delivers check, a fork or mate that a queen cannot.
Take a rook when a queen would stalemate, which is the Saavedra idea.
A bishop is essentially never correct outside composed problems.
And run the two second check on every promotion, because the cost is nothing and the saving is an entire point.