Beating the bots in TAMNA — Easy, Mid and Hard run different playbooks

In the room lobby every bot seat has an Easy / Mid / Hard switch, and only the person who created the room can touch it. Seats you leave alone get a default. The first thing worth knowing is that this default is not fixed.

It flips from Mid to Hard at two wins — but only certain wins count. The game has to have finished, it has to have run at least five turns (games abandoned right after the start do not count), and you have to have been the only human in it. Games you played with friends do not feed this counter, no matter how many you win. If you are not signed in, the default stays at Mid.

The same default is used when Quick play spins up a fresh room and drops three bots into it. Practice mode is a separate world: its three bots are always Easy, and practice games are never recorded, so they never feed the win counter either. You can run the tutorial all day and your default opponent will not move.

A difficulty you pick yourself always wins. In particular, a seat you set to Easy will never be pushed up by your win count. If the default has settled on Hard and you want a relaxed game, lower each seat by hand — and conversely, you can pick Hard from your very first game.

The practical meaning is simple. "The bots suddenly got harder" is not your imagination: from the game right after your second win, the opponent genuinely is a different thing. If you go on a losing run there, it is not that your play got worse. Your opponent picked up more actions — and the rest of this article is a list of exactly which ones.

An Easy bot evaluates nothing. Its opening settlement is picked at random from the legal spots, and its road is picked at random from the empty edges touching that settlement. Dice odds, resource mix, Naru harbours — none of it is read.

The Dokkaebi is the same story. Easy picks any tile other than the one the Dokkaebi is currently on, then robs a random opponent with a building on it. It will not hunt down your 6 or your 8, and it will not think to avoid its own tiles either.

Easy does not trade. It never swaps at the bank and never pitches a deal to you, so when its resources go lopsided they stay lopsided. It never buys or plays development cards. And even when it can afford to build something, it simply ends the turn about 15% of the time.

There are exactly two things Easy does properly. When a 7 forces a discard it sheds its most abundant resource first — that behaviour is identical at every difficulty. And it evaluates trades you offer it using the same rules as every other bot, so the final section of this article applies to Easy too.

This is where the fixed Easy trio in Practice mode matters. What you feel right after the tutorial is not what a real game feels like. Beating Easy is no evidence that you can beat Mid. Above all, the two pressures that define Mid and Hard — trades that will not go through, and a Dokkaebi that parks on your best tile — never appear at Easy. Once the controls feel natural, you learn faster in a room with one Mid seat than in another practice run.

A Mid bot's turn runs in a fixed order: spend any free roads, upgrade to a city, build a settlement, then a 45% chance to end the turn right there, then extend roads, then bank trades, then an offer to a human. Development cards appear nowhere in that list.

What matters is that the 45% fork sits after the building steps. So if the resources are there, cities and settlements always go up and Mid keeps scoring at a normal pace. What vanishes is the trio of road expansion, bank trading and trade offers — and they vanish together, on roughly half of all turns.

Mid picks spots purely on the summed dice odds of the adjacent tiles. Duplicated resources do not lower the score, and a Naru on the corner does not raise it. Its roads have no direction either; they go on the first legal edge. Everything Hard does — weighting resources it does not yet produce, paying up for a 2:1 Naru it can actually use, aiming a road at a good open corner two steps away — is simply absent from Mid.

Three readings follow. First, Longest Route is effectively uncontested against Mid: its roads have no intent, and on nearly half its turns it never reaches the road step at all. Second, Naru corners rarely get contested. A high number may still get taken because of the number, but nobody is racing you to the harbour.

Third, when you see a Mid bot sit on a pile of resources and end its turn doing nothing, nothing is broken. That is the 45% fork, and the resources are still there. Read "it idled this turn" as "it will do two things next turn". Against Mid, calmly locking down your own roads and harbours first is the shortest path to the win.

Before it builds anything, a Hard bot considers exactly one development card to play this turn. The order is fixed: Year of Plenty, Monopoly, General, Road Building. Mid and Easy neither buy nor play them, so this whole section is about Hard alone.

Year of Plenty is used to top up whatever its next build is short of, two cards at a time. If it is short of fewer than two cards, it fills the rest in the order tangerine, basalt, clay, lumber, horsehair. So a turn where Hard plays Year of Plenty is, almost without exception, a turn where something gets built.

Monopoly triggers when every other player combined holds three or more of a single resource. Because it is a combined total, it can fire in a four-player game even when no single opponent has three. Turn that around: the moment you personally stockpile three of one resource, you have pulled the trigger by yourself. On a tie the earlier resource in the order clay, lumber, horsehair, tangerine, basalt is taken. Hoarding basalt and tangerine for a city upgrade is a fine plan — but with a Hard bot at the table, lean towards spending before the pile gets big.

General cards are played the moment Hard holds one. It never sits on them waiting for the perfect Greatest Force timing. The result is that the Dokkaebi moves often in a Hard game and the bot's army count climbs steadily. If you want Greatest Force, budget from the start for the fact that overtaking a bot that started earlier costs you strictly more cards.

One more tell: Hard buys a development card only at the very end of its list. Nothing to upgrade, nothing to settle, no road to extend, no bank trade available, nobody to pitch — only then does it buy. So a Hard bot that suddenly starts buying cards is publicly announcing that it is stuck. In a table of Mid and Easy bots, by contrast, nobody competes for Greatest Force at all; those two points are cheap.

Where you should put the Dokkaebi — on the leader’s 6 or 8, next to whoever is holding the most cards, and never on a tile nobody borders — is general advice that Tamna’s own strategy guide already covers in its section 5 and in that page’s FAQ. What this section is about is not that advice but the formula a bot uses to reach the same decision. And the headline is that Dokkaebi placement is the one area with no difficulty gradient at all: Mid and Hard run the identical formula, and only Easy is random.

The formula: for each candidate tile, weight the adjacent buildings by (city 2, settlement 1) times how often that number comes up, and sum them — except that its own adjacent buildings are subtracted at double that weight. The subtraction is the half that matters in play. A negative double easily cancels out whatever it gained by blocking you, so a bot effectively cannot pick a tile it borders itself. Which means a settlement of yours that shares a single tile with a given bot is not a target for that bot, even if it is your 6 or your 8. The caveat is that this only holds for the bot doing the sharing; the other bots, which do not border that tile, still count it normally. At a table with three bots, price it as "one fewer pair of eyes on that tile," not as safety.

This is also where the hope that "it is only Mid, the Dokkaebi will be soft" dies. Mid buys no development cards at all and skips roads and trading entirely about 45% of the time. For an opponent that cuts that many corners, the Dokkaebi formula is nevertheless character-for-character identical to Hard’s. The Dokkaebi is simply not on the list of pressures you can reduce by lowering the difficulty.

Picking the victim is a separate calculation from picking the tile, and this one is fully deterministic: it robs whichever opponent adjacent to that tile is holding the most cards. There is no draw involved, so as long as you top the count, "I got lucky and it skipped me" does not happen. Leave your best tile exposed while sitting on a fat hand and you will be blocked and robbed in the same move, every time. A bot also always moves the Dokkaebi to a different tile — staying put is excluded from its candidate list — so once a bot’s turn comes round it will come off your tile one way or another, and equally it can come straight back on the next bot turn after you clear it. If the room was created with the Dokkaebi switched off, none of this arises.

Two things to carry into a game. When you choose which settlement to upgrade, weigh not only the strength of the number but which bots border that tile: every bot that borders it is one fewer opponent that will target it. Conversely, the perfect production spot that only you touch is a tile the Dokkaebi will land on every single game at Mid and above. And when a bot discards on a 7 it always sheds its most abundant resource first regardless of difficulty, so a lopsided bot loses the most on a seven — which is exactly why it can be worth waiting out one 7 before you fire your own monopoly.

Whether a bot accepts your offer is decided by one rule set that is identical at every difficulty. Four conditions. It must actually hold the cards you are asking for. You must give at least as many cards as you receive — asking for two and offering one is refused automatically. What you ask for must not cut into what the bot still needs for its cheapest next build. And what you offer must cover a shortfall against the cost of a road, a settlement or a city.

The fourth condition is the one that decides most deals. The costs are: road = 1 clay + 1 lumber; settlement = 1 clay + 1 lumber + 1 horsehair + 1 tangerine; city = 2 tangerine + 3 basalt. So basalt counts as useful whenever the bot holds fewer than three, and tangerine whenever it holds fewer than two. Clay, lumber and horsehair are only ever needed one at a time, so they only register as useful when the bot holds none at all. If you are handing something over, basalt and tangerine are the two cards that get through.

The third condition has a gap in it. A bot only protects the shortfall on its cheapest goal, so once that goal is fully covered there is nothing left to protect. Offer one clay for one tangerine to a bot sitting on 2 tangerine and 3 basalt — a bot one step from a city — and the trade goes through. You can dismantle a fully assembled city before its turn arrives. The moment a bot looks like it is stockpiling is exactly the moment to open negotiations.

If you make an open offer with no specific recipient, the lowest-numbered seat that satisfies the rules takes it first. When you want a particular bot to hand something over, address the offer to that bot.

Offers coming the other way carry a lot of information. A bot only ever pitches to a human, at most once per turn. It asks for one card of whatever its cheapest goal is short of, and pays with one card of the resource it holds most of — and only from a stack of two or more, never its last copy. So the request itself announces the next move: clay or lumber means a road or a settlement is coming, tangerine or basalt means a city is close. Declining costs you nothing, because it will not ask again that turn.