Project story / Warden · 19 Jul 2026
What it took to turn Warden from a brainstorm into a game
Before I made a benchmark that asks AI models to ship games, I tried to make one properly myself. Warden began as a relatively compact idea: a portrait tactics game, built around the satisfying clarity of moving a single piece on a small board.
It became far more work than the brainstorm suggested. That is also what made it useful. A game can be technically complete, visually coherent and still have no reason for the player to make one move rather than another. Warden only became good when I stopped treating the first rules as the idea and started treating them as material to test.
A small board with no excuses
Warden is a deterministic, zero-randomness tactics roguelite. You control one piece on a 6 × 6 dungeon board, descend through ten floors and eventually face the Bone Tyrant. Enemy intentions are visible. There is no lucky damage roll to rescue a poor choice and no hidden result to blame when a run ends.
The portrait, one-thumb format forces the game to stay legible. Movement borrows the immediate literacy of chess, then upgrades mutate what the piece can do across a run. The aim is a sequence of compact decisions where the player can understand the danger, form a plan and feel responsible for the outcome.
The kiting problem
The early versions had a structural flaw: survival was too often about keeping away from everything. If the strongest strategy is to circle the board while enemies shuffle behind you, the game may be solvable but it is not tense. Adding more enemy types would only disguise the problem.
The solution was the Wards. Each floor contains soul-anchors that the enemies march towards. Lose them and the run ends. Suddenly distance has a cost. The player has to interpose, decide which threat matters and sometimes accept danger to protect the objective. The same board and movement rules began producing much better decisions because the player now had something to defend.
The Wards did not add more game around the edges. They gave every move a reason.
Using simulation without designing by spreadsheet
Because the rules are deterministic, Warden can be exercised headlessly across large numbers of boards and builds. The core suite now contains 176 tests. Simulations exposed forced-loss openings, dominant upgrades, difficulty cliffs and rules that sounded fair in prose but behaved differently in combination.
That evidence made it possible to change the design aggressively without guessing about every consequence. It did not answer whether the game felt good in the hand. A bot can tell me that a floor is survivable; it cannot tell me whether spotting the move is satisfying, whether the threat language reads quickly on a phone or whether a ten-floor run has the right rhythm.
The design loop therefore moved between rules, automated simulation and actual play. Some ideas survived. Others were cut even after they had been implemented because they reduced meaningful options or only looked like upgrades. The low-poly art is from a coherent licensed asset set rather than AI-generated imagery; the distinctiveness comes from the rules, composition and presentation rather than pretending generated assets have no provenance.
Where it is now
Warden is launch-staged rather than publicly launched. A signed iOS build, store material and support site exist, but it still needs the remaining Apple account steps and a final real-device TestFlight check. That distinction matters: App Store preparation is not the same as having strangers play it.
Even before that verdict, the project has answered the question that started it. The brainstorm could become a real game, but only after the original idea was challenged repeatedly. The fun arrived late, once I had found the strategy that flattened the experience, changed the stakes and done the less glamorous work of proving the new rules held together.