Citadels: is it smart to take the Assassin?

Why does everybody grab the Assassin?

Because it feels safe. The Assassin acts first, so nobody can kill it, and the Thief isn't allowed to rob it. Leave it in the draft and someone after you might pick it up and name your character. That fear isn't imaginary.

When our bots left the Assassin for their next-best character, an opponent ended up holding it 70% of the time with four players, and the bot that left it was killed that same round 19% of the time. With two players, where you can also throw it away face down, that was 14%.

Freeze the draft, change one pick

A win rate for "players who took the Assassin" would prove little: people take it in different situations than other characters. So we compared like with like. At 152,652 moments in bot games where a player was offered the Assassin, we froze the game and played the rest of it 8 times for every option: take the Assassin and kill as our smart Assassin would, take it and kill nobody, or take one of the other characters on offer. Every option got the same shuffles, so the only difference is that one pick. That came to 8,062,960 replayed games.

"Your next-best pick" below is the character our bot rated highest after the Assassin: what you would have taken if it wasn't there.

With two players the Assassin is the right pick: 56.1% against 49.3%. With four players it is a wash (23.6% against 24.4%), with five a small loss (17.9% against 19.9%).

And that is against the bot's own second choice. Against the best other character on offer, picked on half the replays and measured on the other half so luck can't flatter it, the Assassin trails by 4.1 points with four players and 5.1 with five. With two players it still comes out 3.5 points ahead.

Is it the fear or the kill?

The two options with the Assassin split its value in two. Taking it and killing nobody is pure protection: you are safe, and so is everybody you would have killed. The step from there to taking it and killing is the kill itself.

Holding the Assassin without using it loses at every table size: -9.6 points with two players, -5.4 with four, -4.9 with five. The protection is real, but the character you gave up for it is worth more. Everything the Assassin is worth comes from the kill: +16.4 points with two players, +4.6 with four, +2.9 with five. With more players a kill hurts only one of several opponents, so it stops covering the cost.

With four or five players, when does it pay?

Two situations with four players. When you are ahead, you are the one everybody wants to kill, and the Assassin both protects you and hits whoever is chasing you: +1.4 points. And when you pick last, nobody after you could have taken it anyway, but you still get the kill: +1.1. In the middle of the pack it costs you 1.7 points.

With five players we found no situation where the Assassin clearly beats your next-best pick. The best case is break-even, as the last picker (-0.1).

Look at what else is on offer

The Assassin isn't competing with an average character, it is competing with whatever is still in your hand of options. With four players it is about even against most of the cast, and clearly beaten by one card: the Architect. If both are offered, the Architect wins by 7.6 points (with five players by 7.3). Three districts in one turn is a lot of city.

What if the whole table grabs it?

One frozen pick is not a whole game, and the fear is a habit, not a single decision. So we played full games with one player who takes the Assassin every time it is offered, against bots that pick it on merit (they took it about 36% of the time it was offered with four players). Then the reverse: one player who never takes it, and one merit player at a table where everybody else grabs it. Seats and the crown rotate; 120,000 games per setup with four players.

The grabber gets exactly what it wanted: it is killed in 7% of rounds instead of 11%. It still wins only 21.9% of four-player games, and 14.0% of five-player games against a fair 20%. The pure version of the fear, take it every time and never kill, wins 11.2%. Never taking it at all is fine: 25.5% with four players, 21.4% with five.

Now the twist. At a table where everybody grabs the Assassin, everybody gets killed more often: 14% of rounds per player with four players, against 11% at a table that picks on merit. Nobody leaves it in the draft any more, so it is in play in more rounds, and its kills land more often: 92% of them, against 83%. The one player who picks on merit at that table wins 26.4% with four players and 23.2% with five.

Does the rule of thumb hold up?

The rule card above comes from the frozen picks, so we tested it separately, in new full games: one bot follows the card, the rest pick on merit. It won 76.6% with two players, 26.7% with four and 21.4% with five, against 49.9%, 25.0% and 19.9% for a seat that picks on merit. With four players that also beats both extremes, never taking it (25.5%) and always taking it (21.9%). With five the card says leave it, which is the same as never taking it. At a table of grabbers it held up too: 28.2% with four players.

Two players: take it, or bury it

Citadels for two has its own draft. One character is discarded face down at random, the crowned player takes one of the remaining seven, and from then on each player takes one card and discards one face down before passing the rest. Everybody ends up with two characters. Two things follow. A kill always hits your only opponent, and you can keep the Assassin out of the round without taking it: discard it face down.

So at the second and third pick there are three ways to handle it, and we tested all three against each other. Taking it wins 55.9% of games. Taking your next-best character and burying the Assassin wins 48.6%. Passing it on wins 45.1%: your opponent then kills you in 43% of those rounds. Taking it and killing nobody is the worst of the lot at 39.1%.

The later the pick, the bigger the gain: +5.4 points for the very first pick of the round, +15.3 when it is one of the last two cards. The one exception is the first round (-1.7): nobody has built anything yet, so a kill takes away less. Our merit bot undervalued all this and took the Assassin only 23% of the times it was offered with two players. A player who simply takes it every time beats that bot in 75.8% of games.

Get the game

Everything here uses the original eight characters, which are in every Citadels box: Citadels Classic has just those eight, the current Citadels edition adds alternative characters for every rank. If you mostly play as a couple, Citadels Duel is a separate two-player redesign with a council board, out since October 2026. We have not simulated it, so none of the numbers above apply to it.

Once you hold the Assassin, the next question is who to kill. That one has its own answer: should the Assassin kill the King? And if you are shopping for two, Citadels is in our ranking of the best 2-player board games.

How we tested

We reused the rules engine from our kill-the-King analysis: the eight original characters, the base district deck with the original purple districts, and the current Citadels Classic rules, including the two-player draft with face-down discards. Automated tests check every odd rule, and for this article also that a draft frozen at any pick and replayed with the same choices comes out identical.

The frozen-pick experiment sampled 70,448 two-player, 44,725 four-player and 37,479 five-player picks where the Assassin was on offer, and replayed each option 8 times with identical random numbers (8,062,960 replays). The full games with Assassin habits and the rule-of-thumb test add 3,750,000 games, with seats and the crown rotating. Margins of error are 95% intervals; the frozen-pick ones use the spread between decisions.

The bots draft by valuing each character for their own position and discounting characters likely to be killed or robbed; they pick with some randomness. Their Assassin reads the draft and aims at the turn that hurts most, preferably the leader's. "Picks on merit" means this valuation, which is a heuristic, not the truth: with two players it clearly undervalues the Assassin.

Limits worth knowing. Bots don't bluff, don't talk and don't read faces, and Citadels is a game of all three. A human who picks the Assassin and sells a different story can do better than these numbers; a human whose face says "I have the Assassin" can do worse. Our merit bots also don't learn that a grabber always holds it, so a sharp table could punish grabbers harder than ours did. We tested two, four and five players, not three or six and seven, and not the alternative characters of the current box. None of this is solved play: the rule card is the best of the strategies we tested. More on how we test.