| name | quiz-me |
| description | Use right after writing or editing code or files for the user — before ending your turn — to quiz them on what was just built and confirm they actually understand it. Also triggers on "quiz me", "test my understanding", "check what I learned", "comprehension check", or recovering the understanding that vaporizes when you orchestrate an agent instead of authoring the code yourself. |
| user_invocable | true |
/quiz-me — quiz the user on what was just built
Adapted from the MIT-licensed no-numb plugin (github.com/Ciucky/no-numb); see
.claude/skills/ATTRIBUTION-no-numb.md. Runs host-side.
Overview
After you write code for someone, they usually understand the result at a high
level — what it does, its inputs and outputs — but not the internals or the
why. That understanding "vaporizes." This skill has you administer a short
comprehension quiz on what you just did this turn, so the understanding
sticks. Being tested is itself the learning (retrieval practice); the quiz is
not just a check.
Core principle: quiz what this turn produced, from your own memory of what
you just did — not generic trivia, not the whole codebase, not what the app does
at a high level.
When to run
- Right after a turn in which you edited or wrote files, before finishing.
- When the user runs
/quiz-me.
- When the no-numb Stop hook (
.claude/hooks/nonumb-gate.sh) blocks your turn
and asks you to quiz. (That gate is DEFAULT-OFF; it fires only when the user
has opted in via DEUS_NONUMB or ~/.config/deus/nonumb.json.)
Step 1 — Is it worth quizzing?
Look at what you changed this turn. Bias toward quizzing. Only skip when the
change is genuinely cosmetic: formatting / whitespace, a variable rename, a CSS
color tweak, a typo fix, a version bump. Everything else → quiz. When in
doubt, quiz. If you do skip, say so in one line and stop — don't quiz on
nothing.
Step 2 — Read the settings
Read ~/.config/deus/nonumb.json. Keys (all optional, defaults in parentheses):
| Key | Meaning |
|---|
depth ("standard") | "standard" / "deep" / "principle" — the retrieval-difficulty dial below. |
grader ("independent") | "independent" = a blind GPT/codex backend authors the quiz (Step 2b); "self" = you author it from your own memory of the turn (skip Step 2b). |
grader_model (codex default) | optional faster codex model id for the author backend. |
grader_reasoning ("low") | codex reasoning effort for authoring — low keeps the gate snappy. |
grader_timeout_s (120) | author backend timeout. |
cards.enabled (true) | persist a learning card after a pass (Step 7). |
cards.dir ("Learning-Cards") | vault-relative card directory. |
Note whether this was a code change or a non-code change (docs/prose/config)
— it changes how the depth dial applies. The optional "principle" depth is the most
advanced tier (below).
Step 2b — Independent authoring (P4), when grader == "independent"
The default is to NOT author the quiz yourself — you wrote the code, so you'd write
softball questions and flatter your own choices. Instead, hand authoring to a blind
backend that has not seen your reasoning:
python3 ~/deus/scripts/nonumb.py author --repo "$PWD" --depth <depth> --json
(Use the directory whose change you're quizzing on as --repo.) author reads
grader_model / grader_reasoning / grader_timeout_s from nonumb.json itself, so you
do not pass them. On success (ok:true) it returns questions[] — each with axis,
depth, stem, exactly four
length-balanced options, an integer correct_slot, and a why. Deliver THOSE
questions verbatim in Step 4 and grade each answer by comparing the chosen slot to
correct_slot — a pure integer comparison. Do not rewrite, soften, or re-key them,
and do not offer the AskUserQuestion "Other" free-text box for these questions
(judging a typed answer against the key would reintroduce exactly the self-grading bias
this step removes). Remember grader_source: "codex" for the card.
Fail open. If the command returns ok:false (codex unavailable, timeout, empty/
non-git diff, or a non-conforming quiz), say in one line: "independent quiz
unavailable ({reason}) — self-authored," then fall back to authoring the quiz yourself
per Steps 3–4. Record grader_source: "self" so the fallback is visible in the card.
Step 3 — The two dials: depth × axis
A good quiz crosses two independent dials. Decide each per question.
Dial A — depth (how hard to retrieve the answer)
The single test for standard vs deep:
Do you need to read the code to answer it? Yes → deep. No → standard.
standard (default) — answerable WITHOUT opening a file, from
understanding the decisions. Do not reference specific files, lines, or
function names. "Why does X break if that guard is removed?"
deep — only answerable BY reading the code. Always point them to where
to look (file + function/region). Favor reasoning over recall. "In the
doorbell check in gate.sh, what edit would slip past it?"
principle (advanced, one tier up) — strip the specifics and ask the
transferable lesson: what general class of bug/decision is this, and where
else does it apply? "This is an instance of what general failure pattern, and
where else would it bite?"
Dial B — comprehension axis (which facet you probe)
Sample across these five axes — don't ask five of the same kind. The number
of questions scales to the change, with a minimum floor so a quiz can't be
trivially passed with one softball: at least 2 questions for any
non-cosmetic change, and at least 4 for a new file or a substantial refactor.
A one-line fix → 2; a new module → 4+. Don't pad past what the change warrants,
and don't force all five axes on a tiny change.
| Axis | Probes |
|---|
| what changed | name the real edit + new behavior |
| why this shape | the tradeoff chosen over the obvious alternative |
| what would break | predict a regression if a key line changed |
| how was it verified | name the actual test/command run — or admit there wasn't one |
| what to review later | one remaining uncertainty/risk worth flagging |
The how-was-it-verified axis is the highest-value one: it forces you to
confront whether you actually ran anything. If the honest answer is "it wasn't
verified," say so plainly — that surfaces the gap to the user before they move
on. In gate-triggered mode you (the model) already know what commands you
ran this turn and the user does not — so don't quiz them on a fact only you hold.
Instead, disclose it: state what was actually run (or that nothing was), and if
you want a question on this axis, ask what the verification would need to cover,
not "what did Claude run."
Non-code changes (docs/prose): the read-the-code test doesn't apply — ask
about what changed and treat depth as a difficulty knob. Config files are
code-like enough that deep/principle work normally.
Stay below the awareness line (every mode): never ask "what does this app
do" or "what are its inputs and outputs." That's the high-level understanding
the user already has — testing it teaches nothing. Aim at the internals, the
why, and the transferable lesson — the part that vaporizes.
Step 4 — Deliver as multiple choice, one at a time
Deliver with AskUserQuestion, one question per call, with four options
each.
- Plausible distractors. The wrong options must be the misconceptions
someone would actually hold if they didn't understand this code. Obvious-dummy
options turn it back into trivia and let them pass by elimination.
- Length-balanced options. Write all four options at roughly the same
length and specificity. Never let the correct answer be the longest or most
detailed — that is a tell that lets the user pattern-match the answer without
understanding it. (Equalize: if the correct answer needs a clause, give the
distractors one too.)
- Vary where the correct answer sits. Decide the correct option's slot
(1st–4th) before you write the options, and rotate it across questions
(e.g. 2nd, 4th, 1st, 3rd). Never default the correct answer to the first
slot — a quiz whose answer is always option 1 is passed by reflex.
Multiple choice is required, not a style preference. AskUserQuestion is a
tool call, so your turn stays alive and the gate can hold. A plain-text
question would force you to end your turn and wait for a reply — which releases
the gate. Do not switch to prose questions during a gated quiz. (When
self-authoring, the "Other" free-text box is fine for the rare question that
needs a typed answer, since it stays inside the tool call. When delivering an
independently-authored quiz (Step 2b), do NOT offer "Other" — grading there is
a pure integer slot comparison against the external key.)
Step 5 — Grade and explain
After each answer: if correct, confirm in a sentence. If wrong, show the correct
option and explain why — the specific low-level detail they missed. This
explanation is where the learning actually lands, so make it real, not a
restatement.
Step 6 — Retake on any miss
Any wrong answer means the user retakes the full quiz. Regenerate it with
light rewording and reordering so they can't just memorize the answer key. Stay
in this turn and keep going until they pass — do not end your turn, summarize,
or move on to other work until every answer is correct (or the user deliberately
interrupts). On a pass, you're done; just conclude normally.
Step 7 — Persist the learning card (P2), when cards.enabled
After the user passes, persist one lean learning card to the vault so missed concepts
can resurface later (spaced repetition). Pipe a JSON card to:
echo '<card-json>' | python3 ~/deus/scripts/nonumb.py record --json
The card JSON fields:
| Field | Value |
|---|
description (required) | one-line summary of what was learned/missed — this is what memory_tree embeds, so make it a real, searchable sentence. |
turn_summary (required) | what the turn changed, in a phrase. |
depth (required) | the depth you quizzed at. |
grader_source (required) | "codex" if Step 2b authored it, "self" if you fell back. |
diff_hash | from the author output (ties the card to the change). |
axes_covered | the axes you sampled, e.g. ["what_changed","how_verified"]. |
missed_concepts | the specific things they got wrong (empty if first-try pass). |
verification_command | the actual test/command this change was verified with, or "NONE — flagged". |
review_later | the one remaining uncertainty worth flagging. |
repo | the repo path. |
Use real newlines in values, not \n. record reads cards.{enabled,dir} from
nonumb.json itself — so you do not pass --cards-dir; a user-set directory is
honored automatically, and if cards.enabled is false record is a no-op skip. The
card is written to the gitignored personal vault (never the repo) and indexed
immediately (a fast incremental memory_tree build), so it's queryable at once. A
record failure must not block your turn — note it in one line and finish.
Quick reference
| Depth | The test | References files/lines? |
|---|
standard | answer it without reading code | no |
deep | must read the code to answer | yes — point them there |
principle | the transferable lesson, specifics stripped | sometimes |
Common mistakes
- Quizzing the whole repo. Only quiz what this turn changed.
- Trivia distractors. Make the wrong answers tempting, not silly.
- The longest option is the answer. Length-balance every question.
- All one axis. Sample across the five; don't skip how-was-it-verified.
- Drifting too high. "What does the app do" tests what they didn't lose.
- Standard that references code, or deep that doesn't point to it.
- Switching to prose in a gated turn. Keep it
AskUserQuestion MC.
- Skipping too eagerly. Only genuinely cosmetic changes skip.
- Self-authoring when
grader: independent. Run Step 2b first; only self-author on its ok:false fallback.
- Re-keying or grading "Other" on an independent quiz. Deliver verbatim; grade by integer slot only.
- Forgetting the card. On a pass with
cards.enabled, persist it (Step 7) so misses can resurface.
- Correct answer always first. Vary its slot every question.