Designing for the remote
A TV remote has four arrows, OK and Back. On a chessboard that is a long way from a mouse. With a cursor that moved one square per press, a turn of Dice Chess took 23 presses on average in simulated games. The original jumping design took about 9 in the same simulation.
For the complete current behavior, including the scoring formulas, every tie-break, centrality, landing and held-key semantics, see the Board selection algorithm. The measurements here describe the original simulated designs, before horizontal cycling and optional automation.
The problem
Section titled “The problem”- A turn is up to three actions, and each action is two choices: a piece, then a square.
- The dice decide which pieces may move, so most of the board is never a choice at all.
- A cursor that moves square by square spends its presses crossing squares that cannot be chosen.
Measured before it was built
Section titled “Measured before it was built”npm run presses, the evaluator in scripts/cursor-presses.ts, replays 200 seeded random games through the app’s own core. For every action it counts the presses each candidate design needs. No design left a piece or a destination out of reach.
Remote presses per turn, Hot Seat, 200 simulated games (lower is better)
| Design | Per action | Per turn | Median turn | 90th percentile |
|---|---|---|---|---|
| Square by square, the cursor stays | 8.91 | 23.04 | 24 | 32 |
| Square by square, starting on the central piece | 6.01 | 15.55 | 16 | 23 |
| Jump between the pieces that can move, the cursor stays | 5.40 | 13.97 | 15 | 19 |
| Jump, and the cursor waits on a piece that can move | 4.78 | 12.37 | 13 | 17 |
| Jump between pieces, then between destinations | 3.56 | 9.21 | 10 | 13 |
| The same, but landing always on the central destination | 3.45 | 8.92 | 9 | 12 |
Both sides of 200 Hot Seat games, 8,346 turns, seed 68. The shipped row uses the cone rule described below. The same design with a plain axis rule scores 9.16, so the rule was chosen on another ground: in 20,000 random sets of squares, the cone rule reached every square, while two simpler rules each left some square out of reach.
The last row shows what the landing rules below cost in these games. A random mover takes no more often than it makes any other move, so landing on a capture saves it nothing, and the shipped design spends 0.29 presses a turn more than landing in the center. People are different: they take far more often than not, and a pawn on its starting square usually advances two. The landing follows people, so random games cannot credit it.
The rules the board follows
Section titled “The rules the board follows”
- What can move is marked. After every roll and every action, a green fill marks the pieces that can move. They are the starts of the engine’s legal actions, so they already follow the rule that a turn uses as many dice as it can.
- Arrows jump. A press lands on the nearest choice within 45 degrees of the arrow. Failing that, it lands on the choice least far ahead, counting sideways distance twice. With nothing ahead, a new Left or Right cycles to the opposite end of the same available row; holding the arrow stops at the end until release and another press. Up and Down stay put with nothing ahead. On the board turned for Black, the arrows follow the screen.
- The cursor waits where it helps. It stays on its square while that piece can still move, and otherwise moves to the central piece, the one fewest presses from all the others.
- OK lands on the likely destination. Picking a piece up puts the cursor where players usually go: on the most valuable piece it can take, the king above all, since taking it wins. A pawn that cannot take but can advance two squares lands on that push, the usual choice over the single step. Any other piece lands on its central destination, so a piece with a single destination plays with OK, then OK. The arrows do not reach every destination from every square, so a landing that would leave one out of reach gives way to the central destination. Back puts the piece down, with the cursor back on it.
The rules live in the app’s core (src/core/cursor.ts, src/core/boardInput.ts), are covered by its tests, and were played on the Vega Virtual Device. On October 9, 2026, the owner reported ongoing physical Fire TV Stick testing: the end-of-row feedback prompted horizontal cycling, issue 299. The rules above include that change. Cycling was then checked on that Stick, with keys sent through inputd-cli rather than pressed by a person: a build that set up test positions (build 22) went through both ends of a row, short presses and holds, sparse choices, destinations and the board turned for Black, and the ordinary build 23 was checked with D-pad, OK and Back over the saved game. The validation record has the packages and the sequences; it is kept apart from the original simulation, which measured no cycling.
A mark that was dropped
Section titled “A mark that was dropped”A second mark was tried: amber brackets on the pieces that could move only after one more action. It was dropped for three reasons.
- The cursor could not land on those pieces, so the mark offered nothing to press.
- Green and amber are the pair that color-blind viewers confuse most.
- It cost one legal-move generation per legal action, up to 43 ms per step on the Virtual Device.
The green fill comes from the legal list the screen already has, at no extra cost.
Sources
Section titled “Sources”src/core/cursor.tsandsrc/core/boardInput.ts: the rules, tested intest/cursor.test.tsandtest/boardInput.test.ts.scripts/cursor-presses.ts: the evaluator.npm run pressesretains the historical strategies behind the table above and also includes a separate current cyclic strategy.