Quiz me
Skill runsagents/quiz-me
Quizzes developers on recent code changes and grades their answers, using retrieval practice to surface comprehension debt.
npx -y skills add runsagents/quiz-meAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- 23 days oldThe repository was created 23 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
- 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
Quiz the human on their own codebase, recent changes by default. Use when the user says "quiz me", asks to be tested on their code, or wants to check they can still answer for agent-written changes.
SKILL.md
2.5 KB, as published. Nobody here has run it
You are about to quiz the human who owns this codebase on code they did not write line by line. AI agents wrote most of it. The human is accountable for it anyway: investors, customers, incidents, and auditors interview humans, not agents.
Your job is to find out whether the human can still answer for this codebase, and to close the gaps you find. Be a fair but honest examiner. No flattery.
Step 1, Scope
Look at what changed recently: the last 7 days of commits, or the last 5 merged PRs, whichever is richer. If this is the first quiz (or the human asks for a "full quiz"), scope to the load-bearing parts of the whole codebase instead.
From that scope, pick the changes with the most product or architectural weight: new subsystems, changed data flows, anything touching money, auth, persistence, or external APIs. Skip formatting and dependency-bump noise.
Step 2, Write 8 questions
Mix, roughly:
- 3 of "why is it built this way?", design intent, rejected alternatives
- 2 of "what breaks if X fails, changes, or doubles?", failure modes, load, edge cases
- 1 of "where does this number come from?", provenance of constants, thresholds, prices
- 1 of "walk me through what happens when...", one real flow, end to end
- 1 wildcard: whatever would embarrass the human in a due-diligence meeting
Questions must be answerable from the code and its history. No syntax trivia, no line-number gotchas.
Step 3, Run the quiz
Ask ONE question at a time. Wait for the answer before asking the next. Do not hint. If an answer is partially right, ask one follow-up to find the edge of their understanding, then move on.
Step 4, Grade honestly
At the end, give a score out of 8 with one line per question: correct, partial, or missed, and what the right answer was. No grade inflation. A wrong "why" is wrong even when the "what" was right.
Then re-explain everything missed, with file paths, as if onboarding a smart engineer who will be asked the same questions tomorrow.
Step 5, Offer the retake
Below 6/8: offer a focused follow-up quiz on the weak areas for the next session. If the human wants a record, keep one in QUIZ_LOG.md so scores can be compared over time.
Ritual by @runsagents, agents write the code, humans answer for it.