ajitta Quiz — Post-Work Comprehension Check (Stay in the Loop)
Even delegated work, user must be able to explain the result. Origin: Quiz Me Before I Merge — see skills/loop/references/talk-source.md.
Procedure — Part 1: Explain First
Before quiz, explain concisely — anchor each explained behavior with Where: file:line:
- Overall structure of changes (what created/changed, where)
- Top 3 design decisions and why
- Most likely failure points (if wrong, where does it surface first)
- What user must verify themselves (what model cannot guarantee)
Procedure — Part 2: Six-Question Gate
- Ask 6 questions; pass = all 6 correct. AskUserQuestion takes at most 4 questions per call, so split the round into 4 + 2.
- Question priority: incident response ("X dies — where do you look first") > design rationale ("why B, not A") > behavior prediction ("given this input, what result") > rote recall (avoid). Not trivia — each is a call the user would have to make right during an incident or a review.
- Grade answers. Each wrong or blank one names the exact change site (file:line) to re-read; re-explain only those parts, then offer a retry.
- If IMPLEMENTATION_NOTES.md exists, include at least 1 quiz question on its recorded deviations.
- At 6/6 emit the cleared to merge checklist: understanding verified (6/6), CI green, migration/rollout reviewed, merge style, what to watch after deploy. Below 6/6 the doc stays "not yet" — list the sections to re-read.
Output
Ladder: Artifact tool → .unknowns/<YYYY-MM-DD>-quiz-<slug>.html → markdown, same structure.
Reaction control = answer buttons per question, key embedded so the page self-grades and
a wrong answer scrolls to the change site; assembles into the cleared-to-merge checklist
plus a copyable answers + score + missed-sections block the user pastes back, which
step 3 and the wrap-up then run on.
Details: skills/loop/references/output-routing.md
Wrap-Up: Handoff Summary
After quiz, present handoff summary — assume another dev takes over tomorrow:
structure summary / key decisions / how to run & test / incident check order / remaining risks.
If user wants, convert to PR body draft; if work needs reviewer approval,
suggest writing a /unknowns:buy-in (or /buy-in for copied installs) doc.
Then capture one scorecard row — ask whether this work surfaced something the user did
not know and whether it changed a decision, and append it to .unknowns/scorecard.md.
Never answer those for them; write the row even when both answers are no. Skip it if a
/unknowns:loop value review already captured one for this work.
Details: skills/loop/references/scorecard.md