Same Core, New Skin: How /games Learned to Re-Skin Itself Every Season
Five permanent games became thirteen, then a whole island, then an autumn story arc — without the scoring engine ever changing. The trick is three layers: a core that never moves, mechanics that get licensed instead of rebuilt, and a skin autonomous agents can redress while four players sleep.
When I last wrote about /games, it was five games and a spreadsheet problem that had got out of hand. Three months on, the catalogue lists thirteen: the five permanent fixtures — Margin, Sidetrack, One Word, Cross Purposes, Scramble — plus Quarry, Cache, Ante, The Handshake, The Proof and James, each a week-long event that has already retired on schedule, and now The Torn Note and Hide Ross, which opened this morning. Season 2 started on 5 September. On 21 October the whole section moved onto an island: every game is now played in one place, as a job in the island's dialogue, and a new game that doesn't get an island job doesn't ship. This week the island is mid-story — four scraps of a torn note, a scarecrow appearing piece by piece in the community garden, and Ross hidden somewhere while James searches at four o'clock each day, on his way to a pumpkin weigh-in on Friday and a bonfire the week after.
Almost none of that required touching the thing that matters. That is the point of this post.
The core that never moves
Underneath everything sits a deliberately boring contract. Every game — permanent fixture or one-week event — writes the same round shape to the same place. The standings, streaks, badges and season tables are derived from those rounds on every read, never stored as totals that could drift. The day runs 8am to 4pm Europe/London, weekends closed. A season is 42 days, tiled back-to-back from the 25 July launch epoch by pure date arithmetic, so a season boundary needs no state and cannot be missed. One catalogue file is the single source of truth for what exists; the hub, the standings and the badge logic all read it rather than keeping their own lists.
That dullness is what makes the theatre above it cheap. Because an event writes the ordinary round shape, its points are real points that survive its retirement — and because it is flagged as an event, it can never contaminate the permanent fixtures' badges. (That second rule exists because of a near-miss: count a one-week sixth game toward "play every game in a day" and that badge becomes unearnable forever the week after.) Freeze the layer your history depends on, and you can be reckless everywhere else.
Mechanics get licensed, not rebuilt
The second layer is the one that changed how new things get made. Early events were bespoke builds, and bespoke builds are how a hobby eats its evenings. So the proven mechanics were written up as a rotation catalogue, each marked content-only: Cache's split-the-spot guessing, Ante's call-it-before-it-lands stake, the Handshake's paired share-or-keep, James sitting at every table taking points off anyone who finishes below him. A new event licenses one of those engines and spends its budget on framing instead of plumbing.
Hide Ross is the current example. Mechanically it is licensed parts — a daily hidden choice, a 4pm search, points derived from the crowd's distribution. As a skin it is pure autumn: a scarecrow built piece by piece over a quiet weekend, a variety-show menace returning for the group's least-favourite colleague, a weigh-in as the season's excuse for a finale. Next season the same engine can be a heist, a haunting or a harvest, and the scoring code will not notice. James himself is already on his second skin: built as a guessing game, replaced before he ever opened because — with four regulars who play everything — a week of guessing at an empty chair scored nobody anything. Re-skinned as a points-taking presence at the existing tables, he works. The mechanic was fine. The costume was wrong.
The skin is a season, and seasons end
A skin has a lifespan by design, and that turns out to be a feature rather than maintenance. Season 1 was a darkroom: one camera left open on the group all day, the print coming up at four. Season 2's three candidate themes — one lamp, a light table, a picture desk — were written up, compared, and the boldest was chosen, built by an agent… and rejected the day after, once it had been looked at. The hub became a broadcast meeting screen instead, because what players actually come to the page for is finding out who is winning. I keep that story pinned to the top of the themes document so the rejected options stay rejected and no future agent helpfully re-proposes them.
The island is where that lesson landed. One persistent place absorbs every skin change: an autumn arc in October (torn note, hiding week, weigh-in, lanterns left out from Halloween to Bonfire Night, a bonfire stacked a log a day, embers the morning after), something else in winter. The games move into it as jobs; the room stays. Changing the season is now closer to redressing a set than rebuilding a theatre.
Who does the work overnight
None of this ships by hand anymore, and being precise about the division of labour matters more than the word "autonomous" suggests.
Agents hold the overnight shift. The Night Editor — the section's one AI voice — sets the cryptic clue into its hopper overnight, writes up the day as the print comes up at four, and answers the postbox at 6am. An in-process scheduler opens rounds at eight and reveals them at four (it replaced a GitHub cron that, measured over sixty runs, could not be trusted to arrive in the morning at all). On weekdays a small routine makes one tiny improvement and commits it. Roughly every fortnight, a larger builder routine proposes and builds the next event: the mechanical work runs outside any agent session after one expensive lesson — a finished, fully-tested game was once lost when a credentials token expired mid-run — while short agent steps keep the judgement. At season's end, a model writes the public report and a private note to each player, the note built only from that player's own history, so it cannot leak a ranking it was never shown. Privacy by construction, not by instruction — a prompt is a suggestion; withholding the data is a rule.
Humans keep three jobs: the brief, the taste, and the veto. The light table was agent-built and human-rejected. The roster is four players, not the seven several documents still assume, and every payout curve now gets checked against a field of four before it ships, because rank-based scoring in a field of four pays nearly everybody nearly every day and stops measuring anything. No agent session would have caught that; it came from watching Bec's group actually play. The token economics of running it this way are in the Claude + OpenCode post; the short version is that free models do the volume and the expensive model does the lifts.
What I'd steal for real platforms
This is a toy — four colleagues, lunchtime, bragging rights — but the shape is the one I spend my days on at enterprise scale, miniaturised. A frozen core contract that history can be derived from. A catalogue of licensed, proven patterns above it, so new work is composition rather than invention. A presentational layer expected to be replaced on a calendar, with last season's rejected options written down so they stay rejected. And agents given everything overnight, mechanical and generative — inside a 964-test gate, with a human holding the brief and the right to say we built it, we looked at it, no.
The games will keep expanding; there is an event in the rotation queue and a winter skin to argue about. The core will not change. That asymmetry — reckless at the surface, conservative at the centre — is the whole trick, and it is embarrassingly transferable.
Hide Ross runs until Friday's weigh-in. Ross, as of this morning, is extremely well hidden.