Games

A browser RPG about climbing the corporate ladder at an EPC — engineering, procurement, construction — company, built with Phaser 3 and no build step at all. You play Karsh, freshly promoted to Engineering Manager and working toward Executive: walking the office, fighting in conference rooms and huddle spaces, running site visits across five industries, and negotiating change orders with clients. Opening index.html in a browser is the entire install, and it plays on a phone as well as a laptop — arrow keys or WASD with SPACE on a keyboard, or an on-screen thumb pad alongside SPACE and H buttons that appear on any touchscreen.

The office runs three floors, each with its own palette — GFS for Engineering, CDB for Construction, and the executive floor, where nobody will speak to you until you've made Director and the fights are deliberately far tougher than anything below. Your cubicle is home base and the single hub for every mode: Office, Site Visit, Visitor Day, or PTO. Battles are turn-based and never random — every conference room and utility room holds one battle object you walk up to and engage, against corporate-satire enemies like Scope Creep, Reply-All Storm, and Deadline Wraith. Levelling advances your actual job title up the ladder from Engineering Manager to VP, while Executive has to be earned upstairs: a scripted Performance Review boss fight, then all four executive-floor bosses — the COO, the CFO, General Counsel, and the Board.

Progress deliberately takes both halves of the job. A Bonus Potential reputation score runs alongside XP, and neither a maxed bar nor an XP overflow actually applies until a change order is approved on a site visit — six sites, from a refinery turnaround to a bridge replacement and a water treatment upgrade, each with Safety, Schedule, and Quality checkpoints and a client negotiation at the end. Underneath, there are no image assets at all: every sprite and background is generated procedurally at boot through Phaser's Graphics API, and each character's skin tone, hair, and wardrobe are derived deterministically from a hash of its texture key, so the cast reads as individuals while the game still runs from a plain file:// page. It draws at a fixed 640×480 and scales up to fill the window at 4:3, so it fills a monitor and a phone screen alike — in portrait the game sits above the touch controls, in landscape they tuck into the letterbox bands beside it. The rest is data-driven, which makes new coworkers, sites, and story beats drop-in additions.

A reading adventure for a four-to-five-year-old that is an adventure game first: the child explores a forest, meets characters, finds secrets, and collects treasure, and reading is the magic that makes all of it work. Every word in the world lives in the Great Book of Everything, and one night the words flew out and scattered across the kingdom as little glowing lights — so bridges forgot how to be bridges and doors forgot how to open. The player is a Word Keeper: reading a word calls it home, and the world mends itself a piece at a time. There is no villain and nothing frightening in it. Pip, a very small fox who is very sure he is very brave, comes along as a companion; Luma the owl keeps the Book; and Grumblewink, who took the letters because nobody ever read him a story, ends up a friend.

Chapter 1, the Letter Forest, is fully playable — a thirteen-scene journey that catches letter fireflies, sorts Bramble's mushrooms, follows a butterfly to a secret glade, crosses the Sound River with Poppy, learns spell-building from Luma, and then talks Grumblewink through a five-lock finale mixing every skill, at roughly twenty-three reading interactions across ten to fifteen minutes. Four later chapters — blending, sight words, first sentences, short stories — are sketched in as preview postcards, and nine reusable activities already cover letter recognition, letter sounds, blending, word reading, rhyme, and comprehension. Progress shows as Reading Power with ranks from Letter Explorer up to Word Keeper, never a score or a percentage or a grade, and difficulty adapts per skill invisibly: three clean answers raise the ceiling, two misses quietly lower it, and she is never told either happened.

The design is careful about failure. A wrong answer gets no buzzer, no red X, and no failure screen — just a friendly "Almost!" while the option fades out of the way; the three-step hint ladder gives a picture clue, then the sound, then the answer itself, and hints still earn full Reading Power, so asking for help is never penalised and two misses auto-reveal so she can never be stuck. It autosaves after every single answer. Technically it is as plain as it looks: classic script tags rather than modules so it runs straight off a file:// page, every character and prop drawn as inline SVG, every sound effect synthesised, and the voice coming from the browser's own speech synthesis — each round speaks its cue aloud, never the answer. Nothing leaves the browser: no accounts, no network calls, no analytics, no adverts. A Parent Zone behind a two-and-a-half-second hold and a multiplication question reports per-skill accuracy, adaptive level, letters and words mastered, and hints used, none of it ever visible to the child.

A game for four-to-ten-year-olds who want to be a doctor or a vet, built on one line: people or pets, everyone needs a hero. It opens on a character creator — skin, hair style and colour, glasses, scrubs, lab coat, shoes, accessories, and a “surprise me” die for children who would rather not decide — and then asks which career to start in. Both are always open and swapping between them never costs progress, so a child can run a Doctor case and a Vet case back to back and keep everything they have earned in either. One choice sets the whole difficulty of the game: Little Helper for roughly four-to-six, where the right tool sparkles, arrows point, and a question only ever offers two answers, or Medical Explorer for roughly seven-to-ten, which opens up more tools, more answers, and the real names for things. It is the same game underneath, adapted rather than cut down.

Twenty levels ship — ten Doctor, ten Vet — and each is a patient rather than a stage: a case arrives, the child talks to them, picks tools off the shelf, works out what is wrong and treats it. The step types the engine knows how to play are deliberately varied, covering talking and empathy, choosing a tool and using it on the right part of the patient, reading a result, finding something in an image, putting steps in the right order, scanning, and being shown an explainer, so no two cases feel like the same quiz twice. Helping a patient grows the building: the hospital is an illustrated cutaway whose locked rooms are all visible from the start with their unlock conditions written on them — the X-Ray Room at Level 5, the Laboratory, Animal Recovery, the Emergency Room, and Surgery and the Helipad reserved for finishing a whole track — and each one arrives with a construction animation. Alongside that sit a Doctor Bag of twenty-four collectable tools, each with its own icon, description and explanation of what it actually does, and a Supply Room that spends Hospital Coins on decor, comfy furniture, patient toys, outfits and one buildable new wing. Rewards come in three currencies, and the one that matters is not the obvious one: Hero Stars for solving, Hospital Coins for spending, and Kindness Stars for how the patient was treated on the way.

The design is unusually careful about being wrong. There is no buzzer, no red cross and no failure screen — a wrong tool gets “Hmm… let’s try another tool!” and after two tries the game quietly shows the answer, so a child cannot get stuck and is never told off. Nothing in it is frightening either: no blood, no needles, no emergencies and no death, and the title screen says in as many words that it is make-believe and not medical advice. The accessibility work runs the same direction — every target is at least 56 px, every drag also works as two taps, nothing depends on being able to read because icons, colour, motion and sound all carry the meaning, sound is off-by-default and mutable from every screen, and prefers-reduced-motion drops the ambient animation. Technically it is as plain as the others here: no dependencies and no binary assets at all, with every character, room, X-ray and microscope slide drawn as inline SVG and every sound synthesised. Patients are pure data — a new one is a single object appended to a file, needing no engine changes — and a validator checks each case for tools that don’t exist, hotspots the species doesn’t have, and questions with no correct answer. Progress saves to localStorage behind a versioned, migratable schema, and nothing ever leaves the browser.

A trivia platform whose questions are built from the live internet rather than shipped inside the app. Questions on current events, science, and world geography are generated server-side from cited sources, validated, cached, and served through a serverless API, with scores, friends, leaderboards, and challenges all authoritative on the server. New trivia appears on its own schedule — nobody has to ship a new build for the game to have something new to ask.

Gathering facts and writing questions are kept as separate jobs. Providers fetch real source material — wire services and public broadcasters for news, research-institution feeds for science, REST Countries and Wikidata for geography — and store the title, URL, publisher, and publication date before any of it reaches a model, which is then asked to write questions from that material and nothing else; it is never asked what is true. Geography doesn't trust a model at all: both answers and distractors come out of structured datasets through deterministic templates. Candidates are rejected rather than repaired, against a checklist — four distinct options, no speculative or opinion framing, no relative dates that stop being true once cached, a source URL that was actually supplied, no duplicate fingerprint — and every rejection is recorded with its reason, so a weak batch can be explained instead of guessed at.

The browser never sees the answer key. A question arrives as text and four options; only once the player answers does the server check it, compute the points, record the result, and reveal the correct answer with its explanation and source. Points scale with difficulty and carry a speed bonus that decays to the twenty-second limit, plus streak multipliers — arranged so rapid guessing is never worth it, since a wrong answer costs the streak, an inhumanly fast answer forfeits the bonus, and the client's claimed response time is reconciled against the server's own measurement. Around that sit one-field accounts (a name gets you a player id and a friend code like KARSH-7F4X), friends and leaderboards, head-to-head challenges that freeze both the question ids and the answer order so both players get an identical test, and a global daily challenge seeded deterministically from the UTC date. It is online only, on purpose.