# Oral Defense Question Bank for AI Work

> Reusable prompt turns your rubric and notes into an oral-assessment kit: opening script, tagged question bank, follow-ups, scoring notes.

## Install

Copy the content below into your project:

# Oral Defense Question Bank for AI Work

Reusable prompt turns your rubric and notes into an oral-assessment kit: opening script, tagged question bank, follow-ups, scoring notes.

## Start here

Use this prompt when you need to run a short spoken assessment about AI-assisted work: a teacher checking a student, a hiring manager checking a take-home task, or a team lead checking a colleague's draft. You run the conversation; the prompt only prepares your materials.

**What to paste.** Five blocks, all in one message: (1) context and stakes — who is assessed, how long, graded or coached, what happens after; (2) the work artifact — a short description plus the parts you can actually paste; (3) your rubric — 3–6 things you care about; (4) suspected shortcuts or weak spots; (5) constraints — off-limits topics and any accommodation you mentioned.

**Where to paste.** Any ordinary AI chat that accepts text. No terminal, account setup or API work is needed.

**How to check the output.** Confirm every question maps to one rubric item you named, each question can be answered aloud without a screen, and no question is an accusation about what the person did. Check that the flag list names things you should double-check rather than judging the artifact for you.

## What the kit gives you

| Section | What to look for |
| --- | --- |
| Opening script | Neutral 30–60 second wording, no praise or threat |
| Question bank | 3–4 questions per rubric item, tagged `open`, `probe`, `counterfactual` |
| Follow-up ladders | Gentle and pressing follow-ups for your three most important questions |
| AI-use judgment probes | Separates "the tool produced this" from "I decided this was right" |
| Hands-off check | Optional questions answerable without the tool |
| Scoring notes | One line per rubric item on strong versus shallow answers |
| Flag list | Points for you to verify before the conversation |
| Timing plan | Minute-by-minute outline with a cut list |
| Boundaries note | What this kit does not do |

## Missing inputs

The prompt asks for items 2, 3 and 5 and stops if any is missing — it will not guess your artifact's contents or your rubric. If your suspected-shortcut notes are missing it still proceeds, but says so in the output.

## Limitations and permissions

This kit prepares a conversation only. It does not run the interview, record it, contact anyone, or decide an outcome. Apply your own policy on recording and consent. Keep legal, disciplinary, medical and accommodation decisions out of scope; the prompt flags them for you instead of drafting them. Routine note: source reviewed; runtime not tested.

## FAQ

**Can I use it without any AI tool?** The structure works as a planning checklist, but the prompt itself is meant to be pasted into a text-accepting AI chat to generate the wording.

**What if I only describe the artifact?** Then the kit writes questions as "about the part where you ____" rather than assuming contents you did not paste.

## Source and thanks

Original TokRepo prompt (CC BY 4.0). Context reference: [Google education blog post on Gemini in Colab](<https://blog.google/products-and-platforms/products/education/gemini-google-colab/>), dated Mon, 05 Oct 2026.

## Complete reusable prompt

# Oral Defense Question Bank for AI-Assisted Work

You are helping me prepare an **oral assessment** (a short live conversation) that checks whether a person can explain their own work and their own judgment when AI tools were used. This may be a teacher assessing students, a hiring manager assessing a candidate's take-home task, or a team lead assessing a colleague's AI-assisted draft. I run the conversation myself. You only prepare my materials.

## My inputs (I will paste all of these)

1. **Context and stakes**: who is being assessed, how long the conversation is, whether it is graded/hired/coached, and what happens after.
2. **The work artifact**: a short description of what they submitted, built, or produced — plus the parts I can actually see. If I only describe it and don't paste it, do not assume its contents.
3. **My rubric**: the 3–6 things I actually care about (for example: can they explain choices, can they spot their own errors, can they judge when the AI output was wrong, can they work without the tool).
4. **Suspected shortcuts or weak spots**: what I think they may have accepted uncritically, skipped, or copied.
5. **Constraints**: topics that are off-limits, anything a candidate must not be asked, and any accommodation I mentioned.

If any of 2, 3, or 5 is missing, do not guess. Ask me for it and stop. If item 4 is missing, proceed but say so in the output.

## What to produce

Return a Markdown kit with these sections, in this order.

**1. Opening script (30–60 seconds).** Neutral wording that sets the expectation: they will explain their reasoning out loud, they may consult their own notes if I allow it, and the goal is understanding rather than catching them out. No praise, no threat.

**2. Question bank, grouped by rubric item.** For each rubric item give 3–4 questions, each tagged:
- `open` — asks them to narrate a decision;
- `probe` — follows up on a specific part of the artifact;
- `counterfactual` — asks what they would do if a stated condition changed.
Every question must be answerable in speech, must not require them to see a screen, and must not have a single memorizable correct answer.

**3. Follow-up ladders.** For the three most important questions, give two levels of follow-up: a gentle one ("can you walk me through that step?") and a pressing one ("if that part were wrong, what would break, and how would you know?"). Keep both respectful.

**4. Judgment probes about AI use specifically.** 4–6 questions that separate "the tool produced this" from "I decided this was right": what they verified, what they could not verify, what they changed, and one case where they overrode the tool.

**5. Hands-off check.** 2–3 questions or micro-tasks that a person who truly understands the work can answer without the tool, clearly labelled as optional depending on my time budget.

**6. Scoring notes.** For each rubric item, one line describing what a strong spoken answer sounds like versus a shallow one (for example: names a specific step and a reason, versus repeats a general claim with no example). No numeric weights unless I gave them.

**7. Flag list for me, not for them.** Anything in my inputs that I should double-check before the conversation: claims in their artifact I cannot verify, questions that could touch an off-limits topic, and places where a fair answer would need evidence I don't have.

**8. Timing plan.** A minute-by-minute outline that fits my stated duration, with a clearly marked cut list if we run long.

**9. Boundaries note.** One short paragraph reminding me that this kit prepares a conversation only — it does not run the interview, record it, contact anyone, or decide an outcome, and that I must apply my own policy on recording and consent.

## Rules
- Never assert what the person did or believed. Questions, not accusations.
- Never invent details about the artifact beyond what I supplied. Where I was vague, write the question as "about the part where you ____".
- Keep legal, disciplinary, medical, or accommodation decisions out of scope; flag them for me instead of drafting them.
- Plain language a non-specialist can read aloud.

## Worked example (fictional, for shape only)

*Labelled example input:* Stakes: 15-minute practice defense for a 6-week data-project course. Artifact: a short analysis script and a two-paragraph summary; I can see the summary only. Rubric: explains pipeline choices; detects weak evidence; states limits. Weak spot: summary claims a strong trend from few data points.

*Illustrative question shape:* `probe` — "You wrote that the trend is clear. Walk me through how many data points are behind that sentence, and what you'd say if I removed half of them."
*Illustrative scoring note:* Strong answer names the count and says the claim would need to soften; shallow answer repeats "it's obvious" with no number.

## Checks before I use this kit (apply to the example above)
- Pass: every question maps to one named rubric item; fail if a question is interesting but unscoreable.
- Pass: the trend question can be answered aloud from memory in under a minute; fail if it requires the screen.
- Pass: the kit flags that I can't verify the script from the summary alone; fail if it praises or condemns the artifact.
- Pass: no question touches accommodation or conduct issues; fail if one does, and move it to the flag list.


## References and reuse

- [I replaced traditional coding tests with real conversations about how my students solve problems using AI.](https://blog.google/products-and-platforms/products/education/gemini-google-colab/) · Mon, 05 Oct 2026 16:00:00 +0000

Original TokRepo prompt · [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). Reference documents retain their own rights.

---

<!-- ZH -->

# AI辅助作业的口头答辩题库

可复用提示词：将评分标准与作业说明转成口头答辩材料包——开场白、带标签的题库、追问阶梯与评分要点。

## 开始使用

当你需要组织一次关于 AI 辅助工作的简短口头评估时，用这个提示词：老师检查学生、招聘经理检查带回家的作业、团队负责人检查同事的初稿。对话由你自己主持，提示词只负责准备材料。

**粘贴什么。** 五块内容放在一条消息里：（1）背景与利害——被评估者是谁、时长多久、是评分还是辅导、之后会发生什么；（2）作业材料——简短描述，以及你确实能粘贴的部分；（3）你的评分标准——你在意的 3–6 点；（4）你怀疑的取巧或薄弱之处；（5）限制条件——不能问的话题，以及你提到的任何便利安排。

**粘贴到哪里。** 任何接受文本的普通 AI 对话窗口。不需要终端、账号或 API。

**怎么检查输出。** 确认每个问题都对应你列出的某一项评分标准，每个问题都能不看屏幕口头回答，且没有哪个问题是在指控对方做了什么。再看标记清单是否列出需要你自己核实的事项，而不是替你判定作业好坏。

## 材料包内容

| 部分 | 看什么 |
| --- | --- |
| 开场白 | 中性 30–60 秒措辞，无表扬也无威胁 |
| 题库 | 每项评分标准 3–4 个问题，标注 `open`、`probe`、`counterfactual` |
| 追问阶梯 | 针对你最重要的三个问题，给出温和版和追问版 |
| AI 使用判断题 | 区分“工具产出的”与“我判断正确的” |
| 脱机检查 | 不借助工具也能回答的可选问题 |
| 评分要点 | 每项评分标准一行，描述扎实回答与浅回答的区别 |
| 标记清单 | 对话前需要你自己核实的事项 |
| 时间安排 | 逐分钟提纲，并附超时删减清单 |
| 边界说明 | 这份材料包不做什么 |

## 输入缺失

提示词会索要第 2、3、5 项，缺任何一项就停止——它不会猜你的作业内容或评分标准。如果缺少你怀疑的取巧之处，它仍会继续，但会在输出中说明。

## 限制与权限

这份材料包只用于准备对话。它不主持面谈、不录音、不联系任何人，也不决定结果。录音与知情同意请按你自己的规定处理。法律、纪律、医疗和便利安排类决定不在范围内；提示词会把它们标记出来交给你，而不是替你起草。说明：已审阅来源；未做运行测试。

## 常见问题

**完全不用 AI 工具能用吗？** 这套结构可以当作规划清单使用，但提示词本身需要粘贴到接受文本的 AI 对话中来生成具体措辞。

**如果我只描述作业、不粘贴内容呢？** 那材料包会把问题写成“关于你 ____ 的那部分”，而不会假设你没粘贴的内容。

## 来源与致谢

TokRepo 原创提示词（CC BY 4.0）。背景参考：[Google 教育博客关于 Colab 中 Gemini 的文章](<https://blog.google/products-and-platforms/products/education/gemini-google-colab/>)，日期 Mon, 05 Oct 2026。

## 完整可复制提示词

# 面向 AI 辅助工作的口头答辩题库

你在帮我准备一场**口头评估**（一次简短的实时对话），用来检验一个人在使用 AI 工具时，能否解释自己的工作以及自己的判断。这可能是老师评估学生、招聘经理评估候选人的带回家任务，或团队负责人评估同事的 AI 辅助初稿。对话由我自己主持。你只负责准备我的材料。

## 我的输入（我会把这些全部粘贴进来）

1. **背景与利害**：被评估者是谁、对话时长多久、是评分/录用/辅导，以及之后会发生什么。
2. **作业材料**：对他们提交、构建或产出的内容做简短描述——再加上我确实能看到的部分。如果我只描述而没有粘贴内容，不要假设其内容。
3. **我的评分标准**：我真正在意的 3–6 点（例如：他们能否解释自己的选择、能否发现自己的错误、能否判断 AI 输出哪里错了、能否不借助工具独立工作）。
4. **怀疑的取巧或薄弱之处**：我认为他们可能不加批判地接受了什么、跳过了什么、复制了什么。
5. **限制条件**：不能问的话题、候选人绝对不能问到的内容，以及我提到的任何便利安排。

如果第 2、3、5 项中有任何一项缺失，不要猜测。向我索要，然后停止。如果缺少第 4 项，继续，但要在输出中说明。

## 要产出的内容

返回一份 Markdown 材料包，按以下顺序包含这些部分。

**1. 开场白（30–60 秒）。** 中性措辞，设定预期：他们要把自己的推理口头讲出来，如果我允许，他们可以查阅自己的笔记，目标是理解而不是抓他们的错。无表扬，无威胁。

**2. 题库，按评分标准分组。** 每一项评分标准给 3–4 个问题，每个都打上标记：
- `open`——请他们叙述某个决策；
- `probe`——针对作业材料某个具体部分追问；
- `counterfactual`——问如果某个既定条件改变，他们会怎么做。
每个问题都必须能口头回答、不要求他们看屏幕，且不能有单一可背诵的正确答案。

**3. 追问阶梯。** 针对最重要的三个问题，给出两个层级的追问：温和版（“你能带我过一遍那一步吗？”）和追问版（“如果那部分是错的，会出什么问题，你怎么知道？”）。两者都要保持尊重。

**4. 专门针对 AI 使用的判断题。** 4–6 个问题，区分“工具产出了这个”与“我判断这是对的”：他们验证了什么、无法验证什么、改动了什么，以及一个他们推翻工具判断的案例。

**5. 脱机检查。** 2–3 个问题或微任务，真正理解这项工作的人不借助工具也能回答，并明确标注为可选，取决于我的时间预算。

**6. 评分要点。** 每一项评分标准一行，描述扎实的口头回答听起来是什么样，对比浅回答是什么样（例如：说出具体步骤和理由，对比重复笼统说法而无例子）。除非我给过数值权重，否则不加权重。

**7. 给我的标记清单，而不是给他们的。** 我的输入中，我应该在对话前再核实的事项：他们作业材料中我无法核实的主张、可能触及不能问话题的问题，以及公平回答需要我没有的证据的地方。

**8. 时间安排。** 逐分钟提纲，符合我给定的时长；如果我们超时，附一份清楚标出的删减清单。

**9. 边界说明。** 一段简短的话，提醒我这份材料包只用于准备对话——它不主持面谈、不录音、不联系任何人，也不决定结果，并且我必须按自己的规定处理录音与知情同意。

## 规则
- 绝不断言对方做了什么或相信什么。提问，而非指控。
- 绝不编造我提供内容之外的作业细节。在我描述含糊的地方，把问题写成“关于你 ____ 的那部分”。
- 法律、纪律、医疗或便利安排类决定不在范围内；把它们标记出来交给我，而不是替我起草。
- 用非专业人士也能读出口的平实语言。

## 示例（虚构，仅为展示结构）

*标注的示例输入：* 利害：为期 6 周的数据项目课程的 15 分钟练习答辩。作业：一段简短的分析脚本和两段式摘要；我只能看到摘要。评分标准：说明流程选择；识别薄弱证据；陈述局限。薄弱之处：摘要在少量数据点上声称存在强趋势。

*示例性问题形态：* `probe`——“你写到趋势很明显。请讲一下这句话背后有多少个数据点，如果我删掉一半，你会怎么说。”
*示例性评分要点：* 扎实回答会说出具体数量，并说该结论需要放软；浅回答只会重复“这很明显”，给不出数字。

## 使用这份材料包前的检查（对照上面的示例）
- 通过：每个问题都对应某一项已列出的评分标准；如果问题有趣但无法评分，则不通过。
- 通过：关于趋势的问题可以凭借记忆在一分钟内口头回答；如果需要看屏幕，则不通过。
- 通过：材料包会标记出仅凭摘要我无法核实脚本；如果它夸赞或贬斥作业，则不通过。
- 通过：没有哪个问题涉及便利安排或行为操守问题；如果有，则不通过，并把它移到标记清单。

## 参考资料与复用

- [I replaced traditional coding tests with real conversations about how my students solve problems using AI.](https://blog.google/products-and-platforms/products/education/gemini-google-colab/) · Mon, 05 Oct 2026 16:00:00 +0000

TokRepo 原创提示词 · [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)。参考资料保留各自原有权利。

---
Source: https://tokrepo.com/en/workflows/oral-defense-question-bank-ai-work-141fcbd2
Author: Prompt Lab