you     ›  just fix the import bug
claude  ›  sure. quick question first: why does it report
           500 imported when only 30 rows actually land?
you     ›  ...because it's broken?
claude  ›  (waiting)

That's it. That's the plugin.

Claude keeps the boring half — reading, searching, typing. You keep the half that can't be outsourced: knowing why. Nothing gets written until you convince it you do.

---

🎯 Pop quiz. And it's for you, not Claude.

Real bug. The support ticket says: "the import claims 500 users imported, 0 failed — but only about 30 are actually in the database."

Claude goes digging and surfaces this beauty:

async function importUsers(rows) {
  const failed = [];

  rows.forEach(async (row) => {
    try {
      await db.insertUser(row);
    } catch (err) {
      failed.push(row);
    }
  });

  return { imported: rows.length - failed.length, failed };
}

And then, instead of a diff, it hands you this:

━━━ 🧠 QUIZ ME · round 1 · root cause ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

  Why does it report 500 imported when only ~30 rows land?

  ┌─ Q1/3 cause ────────────────────┬─ importUsers.js:7 ──────────────┐
  │                                 │                                 │
  │  ○ catch swallows the error     │      } catch (err) {            │
  │  ● forEach drops the promise    │        failed.push(row);        │
  │  ○ pool dies on process exit    │      }                          │
  │  ○ push races between callbacks │    });                          │
  │                                 │                                 │
  └─────────────────────────────────┴─────────────────────────────────┘
     ↑↓ to move — the panel shows the code that answer blames

Answer before you scroll. No peeking. You're inside the demo — this is exactly how it feels, minus the mild panic.

👉 It's 3.

Picked 1? Excellent taste in wrong answers. Here's how Claude peels you off it:

If inserts were throwing, failed would be filling up. But the count came back as exactly rows.length — so failed was still empty when the function returned. The callbacks hadn't even run yet. forEach takes the promise, drops it on the floor, and strolls on. importUsers returns a number it invented, and 470 pending inserts die alone in the dark when the process exits.

Get it wrong and nobody just hands you the answer. You get the same idea again — reworded, from a new angle — until it sticks. Get all three right and then the code lands.

Then the receipt arrives:

━━━ 🧠 RECEIPT ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

  ✅  importUsers.js:4   for...of instead of forEach, so each insert is awaited
  ✅  importUsers.js:12  return the failed rows, not just a count, so callers retry
  ⚠️   db.js:22          insertUser takes a client, so the batch shares a transaction

  ▓▓▓▓▓▓▓▓░░░░  2/3 understood — partial on the shared transaction
  🔒 gate re-locked

Every edit named. Nothing bundled, nothing quietly filed under "minor," and no credit for the one you couldn't explain — that one just gets explained to you until you can.

---

⚡ Install

/plugin marketplace add emanzurv/quiz-me-skill
/plugin install quiz-me@quiz-me

Told to Run /reload-plugins to activate.? That's not a suggestion.

Did it land? Type /quiz-me: — you should see quiz-me, review, and status. Nothing? Go check /plugin, someone's fibbing.

It's now on your machine for every project, sitting perfectly still, judging nothing, until you pick a level. 👇

🎚️ Pick your level

	What happens	Cost
1 · On request	Quizzes you when you ask nicely	nothing
2 · One repo	Quizzes you on every real change in that repo	one line
3 · Everywhere	Same, in every project you open	one line
🔒 Hard mode	Claude physically cannot edit until you pass	one small file

Level 1 — say /quiz-me:quiz-me, or just "quiz me on this first". That's the setup. That was it.

Level 2 — drop this into that project's CLAUDE.md:

Before implementing any non-trivial change, use the quiz-me skill.

Ideal for the one codebase that keeps ambushing you at 2am. Your other repos carry on, blissfully ungoverned.

Level 3 — same line, but in ~/.claude/CLAUDE.md. No repo escapes. No takebacks. (Fine: one rm of takebacks.)

Fair warning: levels 2 and 3 are instructions, and instructions have a habit of evaporating around hour three of a long session. Which brings us to…

🔒 Hard mode

Levels 1–3 are a New Year's resolution. Hard mode is a lock on the fridge.

A hook denies Claude's editing tools outright until you've passed, then re-arms itself at the start of every session like a very small, very smug security guard. Off by default, because nobody wants a pop quiz to rename a variable. Arm it for one repo via .claude/quiz-me.json, or for all of them via ~/.claude/quiz-me.json:

{ "enforce": true, "ttlMinutes": 240, "difficulty": "normal" }

• Keys merge, project over global — { "enforce": false } pardons one repo without discarding the rest of your global settings.

• ttlMinutes = how long one pass holds the door open. 0 = forever, you optimist.

• difficulty = easy, normal, or hard. It sets the whole round at once, not just the count:

  	easy	normal	hard
  Questions	2	3	5
  Options	3	4	4
  Distractors	one is obviously wrong	all plausible	all near-misses
  Preview width	5-6 lines	3-4 lines	2 lines
  Getting one wrong	asked again, then moved on	asked until you get it	back to question one

  QUIZ_ME_DIFFICULTY=hard overrides it for one session, for when you want to suffer on purpose. (The skill's Difficulty table is the source of truth if this one ever drifts from it.)

• Your repos stay spotless; the pass marker and the misses log both live over in ~/.claude/quiz-me/, keyed by project path. Nothing quiz-me writes touches your working tree — no stray .claude/ to gitignore, nothing to accidentally commit.

🧊 The surprise exam

Passing a quiz thirty seconds after Claude explained everything isn't memory. It's a rental.

/quiz-me:review 5

Grabs your last 5 commits and quizzes you cold. No diff. No commit message. Just the code, staring back.

Only your commits, and only ones touching files your sessions actually edited — it reads ~/.claude/projects/ to work out which files those are. Someone else's commits on the branch are out of scope; you never wrote them, so there is nothing to retain. Files you have already been quizzed on go to the back of the queue.

3/5 retained — lost the retry backoff, fuzzy on the cache key

Humbling. Extremely the point.

🐉 It keeps score, and it holds a grudge

Every wrong answer gets written down — the concept, and what you actually got wrong — to ~/.claude/quiz-me/, keyed by repo. That log is not a diary. It's a queue.

• Open concepts come back. Next time a change touches one, it gets a guaranteed question — same concept, different angle, tighter distractors than your level would normally hand you. Answer it right and a RESOLVED line closes it out.
• Miss the same thing twice and it becomes a 🐉 boss. Boss questions open the round, always run at hard tightness no matter what level you set, and a wrong answer sends you back to question one. Yes, really.
• 🔥 Streaks count consecutive batches where you got everything right on the first try — every pre-quiz, plus the post-change check. One re-ask, one partial, one override, back to zero. It rides on the round banner and in the hook's own denial message, so a blocked edit tells you what you're defending before a single question appears.

Want the whole picture without triggering anything?

/quiz-me:status

Armed or not, which config file armed it, your level, whether the gate is currently open and for how much longer, your streak, and every open concept with the bosses marked. Reads only — it never quizzes you and never writes a thing.

😴 It's not a jerk about it

Zero quizzes for lookups, formatting, renames, typos, or anything where you already called the bug and the fix yourself. Needing more than one file to explain a change is a strong tell that it's quizzable — but it doesn't invert: a one-file change that turned on a real decision still gets quizzed. Otherwise it shuts up and works.

And when it's 2am and prod is a smoking crater, say "quiz override." It ships. No argument, no justification, no disappointed sigh — just one quiet note that this one went out unverified.

A gate you can't bypass is a gate you uninstall at 2:04am. 🚪

🧯 Kill switch

• Just this once → "quiz override"
• This repo → delete the line from its CLAUDE.md; set { "enforce": false } if hard mode is on
• Salt the earth → /plugin uninstall quiz-me@quiz-me, then delete ~/.claude/quiz-me.json and ~/.claude/quiz-me/

No hard feelings. Mostly.

🩺 Something's off

Symptom	What's actually going on
It didn't quiz me	Your change hit the skip list. Say "quiz me on this first" if you want it anyway.
It quizzed me on something dumb	"quiz override." Onward.
Hard mode isn't blocking anything	Run python3 --version. The hook is Python and fails open on purpose — no python3, no lock, just vibes. Then check enforce is actually true.
It edited before quizzing me	Say "hold — revert that." File goes back, quiz resumes. Hard mode makes this a non-event.
I passed and it's still blocking	Run /quiz-me:status. The pass marker is keyed by repo root, so the usual cause is an expired ttlMinutes, not a lost one.

Requirements: Claude Code. Plus python3 on your PATH for hard mode. That's the entire list.

Hacking on the hook? It fails open on purpose, which means a bug in it looks exactly like a passed quiz — silent. Run the tests:

python3 -m unittest discover -s plugins/quiz-me/hooks

🤔 Why bother

Autocomplete-driven development is gloriously fast — right up until 2am, when something detonates inside code you have genuinely never read. Congratulations: you're now debugging a stranger's work.

The stranger is you. Last Tuesday. 👻

Claude does	You do
The reading	The understanding
The searching	The deciding
The typing	The knowing why

Every "✔️ I reviewed this" checkbox is a lie you tell yourself at speed.

This one can actually fail you. Which is the only reason it's worth anything.

📄 License

MIT — go wild.