739: How to Build a Great Puzzle Game with Josh Collinsworth

Dave Rupert and Chris Coyer introduce Josh Collinsworth as a code, design, and writing “triple threat,” with a possible fourth threat being relentless Balatro playing. Josh says he is gaming less, but Balatro will probably remain his most-played game for a third year because it is familiar and does not require much thought. He has completed the game’s completionist-plus goals twice and moved from Gold Stake to White Stake runs, using a different deck each time. His current streak is only around six because greedy choices and unlucky bosses can still end an apparently easy run.

Josh describes Tupo as a daily browser puzzle built on a four-by-four grid, initially resembling a Sudoku puzzle using numbers one through four in every row and column. A second layer assigns colors, also requiring each color to appear appropriately across the grid. The key third rule is that no number-color combination may repeat: the board must contain one red one, one red two, one blue one, and so on. Players must therefore reason across both layers simultaneously. The puzzle begins easier on Monday, grows harder through the week, moves to five-by-five on Friday, and reaches a particularly difficult Sunday puzzle.

Josh explains that Tupo starts with a complete puzzle having one unique solution, then removes moves while checking that the remaining board still contains a logical deduction. Easy puzzles stop earlier and provide more givens; harder puzzles remove more information and require more deductions involving both colors and numbers. The generator enforces a minimum difficulty but not a maximum, so some Fridays or other days can unexpectedly be much harder. Signed-in-player data shows hundreds of users, while anonymous play is invisible; over time, perfect-game rates have risen and solve times have fallen, though Josh acknowledges that changing difficulty could also affect those trends.

Josh says he was inspired by Clues by Sam and discarded several prototypes, including a textless shape-placement puzzle and a version where all hints were revealed at once. After experimenting with shapes and colors on a grid, he recognized that he was effectively making a two-layer Sudoku and kept numbers as one familiar layer. Because he has color-vision deficiency, he combined colors with shapes for a colorblind mode. Testing easy puzzles with his seven-year-old helped confirm that the concept was approachable. Josh focused on ensuring that puzzles were enjoyable, solvable, varied, and not merely rotations or mirrors of one another.

The hosts emphasize that Tupo’s payoff comes from moving between different kinds of deduction rather than simply filling in a small Sudoku. Josh says the game always includes a point where the player must use the combined color-number rule, with Sunday puzzles requiring several such deductions. Players may be stuck on numbers, switch to colors, and then switch back. Josh often receives messages from players who think a hint is wrong because they have ignored the third layer—for example, a cell might be a two or three, but the two is impossible because the board already contains a blue two.

Players bring familiar Sudoku habits to Tupo, including marking cells with possible numbers, but Josh found that those marks do not help. The number-only and color-only layers are simple enough that the marks add little, while the real deduction requires cross-referencing both layers at once. Recording every possibility merely repeats the same work in a messier form. Josh says players generally inspect which cells can contain a particular number and then cross-reference those cells with a color. He offers an informal tip: when stuck early in a hard puzzle, the color and number appearing most often may point toward the next deduction.

Tupo uses SvelteKit, Tailwind, Netlify hosting, and PocketBase for accounts and puzzle history; PocketBase runs on SQLite and is sufficient for the game’s scale. Weekly puzzle generation runs through GitHub Actions, producing JSON files committed to a repository. The system blocks early access and tries to prevent zero-second solves, though Josh accepts that determined people can work around such protections. He deliberately skipped app stores, partly because web puzzle games are now discoverable through social media and installed web apps, and partly because of Google Play account drama. He wants modest puzzle-pack revenue to sustain the project, not an ad-filled career-making game.

The hosts credit Wordle with popularizing the daily-game format, shareable result statistics, and the expectation that a puzzle can just be a website rather than an app-store download. Chris also connects the format to the pandemic-era appeal of a short, approachable challenge that does not make failure feel like stupidity. Josh wants Tupo to remain simple enough that players do not have to hold too much in their heads, while still offering harder puzzles for serious logic fans. He contrasts this with games such as Balatro and elaborate variant Sudoku, which can demand much more time or specialized knowledge.

Tupo’s hint system acts as a one-step solver: it scans for the simplest available deduction, highlights the relevant cell, and translates the result into readable guidance. Josh expected hints to be common, but signed-in-player data shows that checks are used more often and average hints remain below one per puzzle, even among weaker solvers. Fast players also tend to use the fewest hints. Looking ahead, Josh says the two-layer mutually exclusive format is related to Latin or Euler squares; six-by-six is mathematically impossible in this structure, so the next size would be seven-by-seven, which may require tracking too many colors. Tupo is available at playtupo.com, with extra puzzle packs for purchase.