rehearse-qa

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Rehearse Q&A

问答排练

Drill the Q&A session before it happens. A simulated audience — hostile and curious personas calibrated to the venue — asks questions one at a time, grounded in the user's actual paper and slides; every answer gets coached into a concise, honest, answer-first response; the questions the speaker hopes nobody asks get prepared deliberately instead of dreaded vaguely.

在正式问答环节前开展演练。模拟受众——针对不同场景校准的尖锐或好奇角色——会基于用户的真实论文和幻灯片逐一提问；每次回答都会被指导修改为简洁、诚实、先给出答案的回应方式；演讲者“希望没人问”的问题会被刻意准备，而非模糊地畏惧。

When to use

使用场景

"Grill me on my paper" / "rehearse the Q&A for my talk" / "mock Q&A"
"What will the audience ask?" / "what's the worst question I could get?"
"Practice my thesis defense / viva" / "prep me for job-talk questions"
Poster-session prep (continuous Q&A, 2-minute and 5-minute pitches)
After
```
write-talk-script
```
/
```
make-slides
```
— the talk is built, now the unscripted part gets rehearsed. After
```
simulate-reviewers
```
— its weakness list seeds the dreaded-question inventory.

“针对我的论文对我提问” / “为我的演讲排练问答环节” / “模拟问答”
“观众会问什么问题？” / “我可能遇到的最糟糕的问题是什么？”
“练习我的论文答辩 / viva” / “为求职演讲的提问做准备”
海报展示环节准备（持续问答、2分钟及5分钟演讲）
在
```
write-talk-script
```
/
```
make-slides
```
之后——演讲内容已完成，现在需要排练非脚本化的部分。在
```
simulate-reviewers
```
之后——其列出的薄弱点可作为“畏惧问题”清单的基础。

Inputs

输入信息

The paper (and slides/script if they exist), in any readable form. Process them transiently — never copy paper text into this repo.
The setting and slot: conference talk / lightning / keynote / poster / defense / job talk, plus the Q&A length in minutes (ask if unknown).
Optional but better: a venue profile
```
venues/conferences/<venue>-<year>.yml
```
(schema in
```
venues/schema.yml
```
) so the audience matches the venue family. No profile?
```
parse-cfp
```
can create one, or run with the generic audience.

论文（以及幻灯片/演讲稿，如果有的话），任何可读格式均可。仅临时处理——绝不将论文文本复制到本仓库中。
场景与时长：学术会议报告 / 闪电演讲 / 主旨演讲 / 海报展示 / 论文答辩 / 求职演讲，加上问答环节的分钟数（若未知则询问用户）。
可选但推荐：场景配置文件
```
venues/conferences/<venue>-<year>.yml
```
（配置文件格式见
```
venues/schema.yml
```
），以便模拟受众匹配该场景类型。没有配置文件？
```
parse-cfp
```
可生成一个，或使用通用受众进行演练。

Process

流程

Build the drill plan. Run:
```
python3 scripts/qa_drill.py --setting conference-talk --minutes 3 \
    --venue venues/conferences/<venue>-<year>.yml
```
Deterministic and offline. Emits the slot math (how many questions the live slot actually fits, how many to drill), the persona lineup with per-persona quotas (venue-family calibrated when
```
--venue
```
is given), the round plan, answer-time targets, and a transcript skeleton for step 6.
```
--json
```
for machine output;
```
--help
```
for all settings. Exit codes: 0 ok, 2 bad arguments or missing/unparsable profile.
Re-verify the slot — mandatory. Venue profiles do NOT store talk slots, and Q&A lengths change per year, track, and session. Check the venue's live presenter instructions (start from the profile's
```
cfp_url
```
and
```
website
```
) for slot length, Q&A minutes, and format (chaired Q&A, no Q&A for lightning, poster logistics). Rehearsing to the wrong clock trains the wrong answers; state what was verified and when.
Read the paper and build the dreaded-question inventory. Mine limitations, claims, experimental scope, assumptions, cut material, rebuttal history, odd numbers, ethics/data provenance — full checklist and the per-question prep-card template in references/dreaded-questions.md. Rank 8–15 questions by probability x damage and build an honest answer card for each. If a weakness is fixable before the talk, say so — fix beats rehearsal.
Run the drill, one question at a time. Follow the round plan (warm-up → hostile gauntlet → dreaded finale → rapid-fire → curveballs). Ask in persona, grounded in the actual paper/slides, then WAIT for the user's answer before continuing — never dump a question list. Persona voices, follow-up behavior, and venue/setting calibration are in references/audience-personas.md. Prior-work rule: a drill question may only cite real papers verified via
```
find-papers
```
+
```
verify-citations
```
, otherwise it stays nameless ("suppose someone claims prior work did X"). Never invent a citation.
Coach every answer. After each user answer, break persona and give the five-part feedback (verdict / what worked / the one fix / a model answer built only from what the paper supports / re-drill if it failed) per references/answer-coaching.md. Coach the answer-first template: headline sentence, one piece of evidence, stop. Re-ask hard-failed questions later — an answer is drilled only when it lands twice.
Grade the transcript deterministically. Record the exchanges in the skeleton from step 1 (the user's answers as spoken/typed), then run:
```
python3 scripts/grade_answers.py transcript.md --target 45 --max 75
```
(Targets come from the drill plan.) Flags per answer: estimated speaking time vs target/cap, unanswered questions, hedge openers, filler density, and buried answers to yes/no questions. Exit codes: 0 ready, 1 re-drill needed, 2 bad input.
```
--json
```
for machine output.
Deliver the readiness report. Summarize: questions that land, questions needing another round (with the one fix each), the dreaded- question crib sheet (question → memorized headline → one number), and any "fix the slide instead" items routed back to
```
make-slides
```
/
```
write-talk-script
```
.

制定演练计划。运行：
```
python3 scripts/qa_drill.py --setting conference-talk --minutes 3 \
    --venue venues/conferences/<venue>-<year>.yml
```
该过程是确定性且离线的。输出内容包括：环节计算（实际问答环节可容纳的问题数量、需要演练的问题数量）、带角色配额的受众阵容（若指定
```
--venue
```
则会根据场景类型校准）、轮次计划、回答时长目标，以及步骤6所需的转录稿框架。使用
```
--json
```
可生成机器可读输出；使用
```
--help
```
可查看所有设置。退出码：0表示正常，2表示参数错误或配置文件缺失/无法解析。
重新确认环节时长——必填项。场景配置文件不会存储演讲环节时长，且问答时长每年、每场、每个会议环节都可能变化。请查看场景的实时演讲者指南（从配置文件的
```
cfp_url
```
和
```
website
```
进入）确认环节时长、问答分钟数及形式（主持式问答、闪电演讲无问答、海报展示流程）。针对错误时长进行演练会导致训练出不合适的回答；请记录确认的内容及时间。
阅读论文并构建“畏惧问题”清单。挖掘论文中的局限性、论点、实验范围、假设、删减内容、反驳历史、异常数据、伦理/数据来源——完整清单及单问题准备卡模板见references/dreaded-questions.md。按“概率×影响”排序选出8-15个问题，并为每个问题制作诚实的回答卡。若某个薄弱点可在演讲前修复，请告知用户——修复优于排练。
开展演练，逐一提问。遵循轮次计划（热身→尖锐提问挑战→畏惧问题终局→快速问答→突发问题）。以对应角色的口吻提问，基于真实论文/幻灯片，然后等待用户回答后再继续——绝不直接抛出问题列表。角色语气、跟进提问方式及场景校准规则见references/audience-personas.md。已有研究引用规则：演练提问仅可引用经
```
find-papers
```
+
```
verify-citations
```
验证的真实论文，否则需匿名表述（“假设有人声称已有研究做了X”）。绝不编造引用内容。
指导每一个回答。用户回答后，切换出角色身份，按照references/answer-coaching.md给出五部分反馈（评价/亮点/核心改进点/仅基于论文内容的示范回答/若未达标则重新演练）。指导用户采用“先给出答案”的模板：核心句、一项证据、停止。未达标的难题稍后重新提问——只有当回答连续两次达标时，才算完成该问题的演练。
客观评分转录稿。将问答内容记录到步骤1生成的框架中（用户的口头/书面回答），然后运行：
```
python3 scripts/grade_answers.py transcript.md --target 45 --max 75
```
（时长目标来自演练计划。）针对每个回答标记：预估发言时长与目标/上限的对比、未回答的问题、模糊开头、填充词密度，以及对是非题未直接作答的情况。退出码：0表示准备就绪，1表示需要重新演练，2表示输入错误。使用
```
--json
```
可生成机器可读输出。
交付准备就绪报告。总结内容：已达标的问题、需要再次演练的问题（每个问题对应一项核心改进点）、“畏惧问题”速查表（问题→记忆核心句→一项数据），以及任何需反馈至
```
make-slides
```
/
```
write-talk-script
```
的“修改幻灯片”建议。

Output

输出内容

An interactive drill session plus, at the end (in chat; written to a file only if the user asks): the graded transcript report from

grade_answers.py

, the dreaded-question crib sheet for morning-of review, and the re-drill list. No predictions — a drilled answer is preparation, not a guarantee of what gets asked.

一场交互式演练环节，结束后（在聊天中；仅当用户要求时才写入文件）：来自

grade_answers.py

的转录稿评分报告、供演讲当天早晨复习的“畏惧问题”速查表，以及需重新演练的问题列表。不做预测——演练后的回答是准备，而非对实际提问的保证。

Adapt to your discipline

适配你的学科

Audience lineups are keyed on the venue

family:

field in

scripts/qa_drill.py

(

FAMILY_LINEUPS

) — fork and add your community's audience (e.g. a humanities seminar respondent, a clinical grand-rounds panel) plus any new personas in

PERSONAS

and references/audience-personas.md. The settings table (

SETTINGS

) takes new formats the same way.

受众阵容由

scripts/qa_drill.py

中的场景

family:

字段（

FAMILY_LINEUPS

）决定——你可以复刻代码并添加所在领域的受众（例如人文研讨会的提问者、临床大查房的评审团），同时在

PERSONAS

和references/audience-personas.md中添加新角色。设置表（

SETTINGS

）也可采用相同方式添加新场景格式。

Guardrails

约束规则

Honest answers only. Never coach wording that hides, minimizes, or misrepresents a limitation or result; decline "help me avoid admitting X" framings and offer the concede-and-scope answer instead (it also performs better). Fixable weaknesses should be fixed, not rehearsed around.
Never fabricate citations — in questions or model answers. Prior-work references in drills go through
```
find-papers
```
+
```
verify-citations
```
or stay nameless.
Never put claims in the user's mouth the paper cannot support; model answers use only the paper's own evidence.
Re-verify slot/session facts against the live venue pages (step 2 is not optional); profiles never carry talk-slot ground truth.
Process the paper transiently; never store paper text in this repo.
Never contact session chairs, committees, or any submission/conference system on the user's behalf.
This is rehearsal, not prophecy: never claim the drilled questions are what will actually be asked, and never predict talk reception.

仅允许诚实回答。绝不指导用户使用隐瞒、淡化或歪曲局限性/结果的表述；拒绝“帮我避免承认X”这类请求，转而建议采用“承认并限定范围”的回答方式（这种方式效果更好）。可修复的薄弱点应优先修复，而非通过排练规避。
绝不编造引用内容——无论是在提问还是示范回答中。演练中的已有研究引用必须经过
```
find-papers
```
+
```
verify-citations
```
验证，否则需匿名表述。
绝不将论文无法支持的观点强加给用户；示范回答仅使用论文自身的证据。
必须根据场景实时页面重新确认环节/会议信息（步骤2为必填项）；配置文件绝不会包含演讲环节的真实信息。
仅临时处理论文；绝不将论文文本存储在本仓库中。
绝不代表用户联系会议环节主席、委员会或任何投稿/会议系统。
这是排练，而非预言：绝不声称演练的问题就是实际会被问到的问题，也绝不预测演讲的反响。