quiz-me Skill Interview Log
Assumptions Checkpoint
| # | Assumption | Tag | Status |
|---|---|---|---|
| 1 | Reads alignment HTML pages and quizzes user on contents | [from request] | Confirmed |
| 2 | Input is a file path, e.g. /quiz-me docs/proposal.html | [inferred] | Confirmed |
| 3 | Questions test relationships, not surface facts | [inferred] | Confirmed |
| 4 | One question at a time | [inferred] | Confirmed |
| 5 | Grade/evaluate answers, push back on vague | [inferred] | Confirmed |
| 6 | No file output — conversation only | [inferred] | Confirmed |
| 7 | Local skill (~/.claude/skills/) | [inferred] | Corrected → repo-managed, so users can quiz themselves about their own codebases |
Interview Questions & Responses
Turn 1: Source Identification + Overlap Check
Q1: What is "the alignment page" you want users quizzed on?
- Options: Anthropic alignment page / doc in repo / configurable URL
- Answer: Alignment page is an HTML file in this or any repo, created by most skills, describing the agent's understanding of what the user wants. Also noted: "Grill-me should not be a skill or a pack, is it in our skills?"
Q2: How does quiz-me differ from the existing taste-calibration skill?
- Options: Different purpose / didn't know about taste-calibration / complement
- Answer: Didn't know about taste-calibration
Deviation: User was unaware of taste-calibration. Clarified that taste-calibration builds alignment for new work (Socratic questioning to shape intent), while quiz-me tests comprehension of existing docs. Different purposes confirmed in subsequent turns.
Turn 2: Assumptions + Strictness
Q1: Assumptions checkpoint (7 assumptions listed)
- Answer: Corrected assumption 7 — should be repo-managed, not local-only, so users can quiz themselves about their own codebases
Q2: What should happen when the user gives shallow answers?
- Options: Adversarial/strict / Educational/firm / Adaptive
- Answer: Adversarial / strict
Turn 3: Input Scope, Termination, Output
Q1: Should the skill accept any readable file or only HTML alignment pages?
- Options: Any readable file / HTML only / HTML + markdown
- Answer: "Let the user choose, all by default" — interpreted as: accepts any readable file, defaults to finding all alignment/spec files
Q2: How should the quiz end?
- Options: Until proven / Fixed count / Section-complete
- Answer: Until proven
Q3: Should the skill produce a verdict?
- Options: Verdict + summary / No verdict / Score card
- Answer: Yes — verdict + summary (pass/fail with weak areas noted)
Turn 4: Coverage Checkpoint
Summary presented: Full skill contract covering name, location, input, trigger, behavior, strictness, termination, output, side effects, commit policy, and benchmark approach.
- Answer: Looks complete — proceed to write the brief
Deviations from Initial Request
- Scope expanded: Original request mentioned "the alignment page" (singular); final contract covers any readable document with a default of finding all alignment/spec files
- Location changed: Initially assumed local skill; corrected to repo-managed global skill
- taste-calibration discovered: User was unaware of existing taste-calibration skill; confirmed quiz-me serves a different purpose (comprehension testing vs. alignment building)