Dead Keys 復活經典打字遊戲,Kiro 承擔主要開發工作
Dead Keys: I Brought Back the Typing Games. Kiro Did the Heavy Lifting.
Dead Keys 重現經典街機打字遊戲,會記錄用戶常按錯的按鍵、以熱力圖呈現,並據此生成練習。Kiro按規格先行方式打造核心功能,依據19份規格,在約35小時內完成開發。遊戲可直接遊玩,無須帳戶;AI 練習功能由 Amazon Nova Micro 提供,預設關閉。
The keys you miss become the lesson.
TL;DR: Dead Keys is my entry for the Kiroween 2026 hackathon, and this post is written as part of it. It rebuilds the 1980s arcade typing games as a keyboard trainer: it counts every key you miss, paints them on a heat map, and builds your next drill around them. Kiro built it spec first, from 19 specs, with hooks, two custom agents, an MCP server and 950 passing tests, in about 35 hours over three days.
I learned to type on a typewriter in middle school. The typing games on our Mac at home are what made me better. A word came at you, and you typed it in time or you didn't. That was the whole lesson, and it worked.
Then I went looking for a number. The National Center for Education Statistics tracks the courses US high school graduates take:
| Year | Grads who earned keyboarding credit |
|---|---|
| 2000 | 44.1% |
| 2019 | 2.5% |
Typing classes mostly went away. The keyboard didn't. Kiroween has a Resurrection category, "bring your favorite dead technology back to life," so I brought the games back.
What it does
You can play it right now at dead-keys.vercel.app. No account, and your progress stays in your browser.
Leave it alone and it plays itself.
| Part | What happens |
|---|---|
| Two classic modes | Words fall from the top, or close in from four corners, the way the old games did it |
| The heat map | Your keyboard on screen, each key colored by how often you miss it |
| Drills | One button builds a drill weighted toward your red keys |
| 14 lessons | Home row first. You pass on accuracy, not speed |
| Extras | A daily drill with the same words for everyone, a ghost of your best run to race, a Spanish word pack, and offline play once it's loaded |

Home row first. Lesson 4 is next because of the misses on t and r.

Falling words. Type them before they land.

The heat map. This is sample progress from three scripted warm-up rounds.
Miss a key, see it, drill it. The two-minute demo shows the whole loop:
What the model doesn't get to do
Dead Keys has an optional AI layer. Amazon Nova Micro on Bedrock can write practice lines aimed at the keys you miss. It's off by default, and the game is fully playable without it.

Every drill says who wrote it. This one is built in.
The code decides which keys are weak, from counts. The code checks every line that comes back: the right characters, the right length, and every word has to be a real English word from a dictionary list. If the model times out, goes over the daily cap, or sends back something odd, the game serves a built-in drill. Every drill says which one you got.
Built spec first with Kiro
Kiro built the core from the specs: the engine, storage, the first screens, gameplay, the drills and the MCP server. It also ran the commits, pushes and deploys from its own terminal. Claude (Opus 5.5) wrote the specs, reviewed each block, and built some of the later features to save Kiro credits.
My favorite part of Kiro is spec-driven development. It keeps me organized. Every block starts as a requirements file, then a design, then a task list with one file per task. I always know where the build is, and so does the agent.

The .kiro folder, counted straight from the repo.
The numbers, read from the repo the day I recorded the demo:
| Count | |
|---|---|
| Specs | 19 |
| Requirement lines (WHEN ... SHALL) | 279 |
| Tasks done | 165 |
| Hooks (4 active, 1 retired) | 5 |
| Unit and property tests passing | 949 |
| axe accessibility scans, issues found | 15 scans, 0 issues |
The spec said "flag, don't guess"
In Block 1, Kiro wrote the data shapes, the contract that everything else builds against. The spec told it to stop and ask instead of inventing. It came back with ten gaps. What format are timestamps? Can an old save with schema version 0 still load? What stops someone storing a hundred ghosts and filling up the browser's storage? Should the AI be called when there are no weak keys?
Ten small questions, about five minutes of answers, and no rework later. Those questions would have come up without a spec too, just later, as bugs.
The hook that never fired
Hooks run on their own when something happens. The line-cap hook keeps every source file under 150 lines, so no file grows into a monster. We planned it to run after each task.
It never ran. The trigger we picked fires when Kiro's task runner finishes a task, but we were ticking tasks off by editing the task file, so that event never happened. We checked the trigger names against Kiro's hook docs, moved it to fire when the agent finishes a turn, and then proved it. A temporary marker wrote one timestamp per run, and I saw "Run Command Hook line-cap" in the panel myself. Then the marker came out.
Since then, a hook doesn't count until we've seen it run. That day I also found the linter running on defaults across the whole repo, because nobody had written its config file. 47 errors, fixed the same day.
A property test caught what examples wouldn't
Property-based tests throw hundreds of generated inputs at a function instead of a few hand-picked examples. In the engine, one found that the words-per-minute math returned NaN for a tiny enough time, because one step of the division underflowed to zero. None of the example tests used a time that small. Kiro rearranged the math and the test went green.
We also broke a rule on purpose, to prove a property test could fail. It did, with a counterexample, and the rule went back in.
An MCP server for the game's own checks
In Block 15, Kiro built a small MCP server for the repo with two tools. One validates a word pack (1,500 English items and 1,421 Spanish items, both OK). The other runs the AI's golden test set, and only when you pass confirm: true, because that one costs money. It refuses paths outside the packs folder, and a test checks that no secrets or drill text come back.
The first thing it got wrong was ours
Block 6 was the first real screens. Both reviewers recommended failing it, for a title-screen keyboard bug and some accessibility and test gaps. I failed it, but the bigger problem was one no check caught: the screens didn't look like the design we'd locked.
Kiro hadn't done anything wrong. The spec listed the design mock as a "reference," and no task said to match it, so Kiro built a plainer version, and nothing in the checks measured the difference. The gap was in the spec Claude wrote.
It's the second time I've hit this. Last month, on the invoicing app I built for one person, eight slices and an end-to-end test all passed against a plan that was wrong, and I only caught it by using the app myself. What I took from both: passing checks tell me the code matches the plan, not that the plan is right. So this time the fix went into the checks. Each screen's screenshot has to match its mock, side by side, approved by me. The rebuilt screens are the ones in the video.
Two more surprises
The first deploy from Kiro's terminal went straight to production. On a brand-new project, the Vercel CLI made the first deployment production even without the production flag. Nothing secret was in it, so I left the site up, and now every deploy runs a build check first and stays a preview unless I say otherwise.
And the AI connects to AWS without a single access key. The function uses Vercel's OIDC federation to assume an AWS role made for the drill route. There's no key to leak, and the AI drills have a daily cap in the code, because a budget alert only tells you after you've spent the money.
I ran my own tools on it
I build tools that check things, so the last step before submitting was pointing two of them at Dead Keys.
Agentis Lux checks how a site reads to an AI agent that fetches the raw page. Dead Keys scored 97 out of 100. It caught two things: no skip-to-content link near the top, which keyboard users miss too, and two external links without a rel attribute.
README Clew 2 checks a README's claims against the code. At commit fd6bd7c, nothing in the README was contradicted, 6 claims were verified, and 73 were the kind code can't check, like my typewriter story. The fair hit: six dependencies the code uses that the README never names, including the MCP SDK and the AWS SDK.
None of it breaks the game, so all three went on the v2 list.
What it can't do
- It needs a physical keyboard, and only QWERTY layouts for now.
- Gameplay is visual. Menus, lessons, results and progress work with a screen reader, and an untimed practice mode is the accessible way to practice.
- The AI caps are approximate, because each server instance counts on its own.
- No letter hotkeys yet. Arrow keys, Enter and Escape work on the menus, but a tester on the Kiro Discord didn't find that, and suggested letter hotkeys like P for Play and a quicker sound switch. Both are on the v2 list.
Try it
Play at dead-keys.vercel.app. The code is MIT licensed:
earlgreyhot1701D
/
dead-keys
A typing game, back from the dead. The 80s typing games MasterType and Type Attack, rebuilt as a real trainer for anyone starting or rebuilding keyboard skills. Accuracy-first lessons, a heat map of your missed keys, adaptive drills, and progress saved only in your browser. Built with Kiro for Kiroween 2026.
Dead Keys
A typing game, back from the dead.
I learned on a typewriter in middle school, and the Mac games at home helped me improve. In 2000, almost half of US high school grads took a typing class. By 2019 it was 2.5%. So I brought the games back.
The number: US high school graduates who earned keyboarding credit fell from 44.1% in 2000 to 2.5% in 2019 (NCES, NAEP High School Transcript Study, Table 1).
Play it: dead-keys.vercel.app
Dead Keys rebuilds the 1980s typing games MasterType and Type Attack as a keyboard trainer that notices which keys you miss and drills them. Built with Kiro for the Kiroween Hackathon 2026, category Resurrection.
Two resurrections
MasterType (1981) and Type Attack (1982). Words fall toward you, and you type them before they land. Dead Keys keeps that pressure and the arcade feel, with synthesized sound and a…
Give it a few rounds and the heat map shows you exactly which keys you've been dodging. In the demo they're t and r.
If something breaks or feels off, this short form goes straight to my v2 list. Two playtesters and my own code review all got stuck on the first menu, so I made one fix before judging: when the menu opens, Enter now starts Play. Everything else on the list lands after Kiroween.
Did a typing game teach you to type, or did you never get taught at all? I'd love to hear which one.
Quick context if you're new here. I work in the California courts, running court operations for the county. I started building with AI in July 2025 and have been learning in public since. I build the Clew Suite, civic tech tools that make complex systems easier to inspect.
Written in November 2026 from the repo, its build log and the public docs linked above. These tools move fast, so check the source before you quote me.
AI Assisted. Human Approved. Powered by NLP.
來源:DEV Community · dev.to

