第二課：把研究問題寫成一句可回答的問題Turning Your Question into One Answerable Sentence
為什麼這件事會出錯Why this goes wrong
有人在 askLLM 裡打了「幫我分析這個資料」，勾選 Submit，收到一段內容豐富、 但完全不對題的回覆——它猜了一個描述統計加一個相關分析，剛好都不是使用者真正 想問的問題。使用者接著在同一個問題欄後面補一句「不對，我是想問兩組有沒有差 異」，再次勾選 Submit。可是 askLLM 沒有對話記憶：第二次送出時，AI 完全 不知道「不對」是在回應什麼，只會把整段文字當成一個全新、獨立的問題重新回答。 模糊的指令加上單輪限制，結果就是來回猜測、浪費呼叫次數，答案還是對不上。
Someone typed “analyze this for me” into askLLM, ticked Submit, and got a long, confident reply. It guessed at a description plus a correlation. Neither was the question the user actually had in mind. The user then appended “no, I meant whether the two groups differ” to the same question box and ticked Submit again. But askLLM has no conversation memory. On the second call the AI has no idea what “no” is responding to. It simply re-answers the whole text as a brand-new, standalone question. A vague instruction plus a single-turn design produces guesswork, wasted calls, and an answer still off the point.
askLLM 是單輪問答askLLM is single-turn
每次勾選 Submit 都是一次獨立、完整的呼叫；沒有前一輪的記憶可以延續。askLLM 內建的防抖機制只在「這次的問題、摘要、persona、語言、system prompt 完全跟 上一次相同」時才會用快取結果，而不是把兩次問題串成對話。應用內建的操作提示 也明講：修改問題前要先取消 Submit 勾選，改好再重新勾選——因為每次改動都可能 觸發新的呼叫。這代表：提示詞必須自足，把背景、變項、比較方向一次寫完， 不能指望「下一句再補充」。
Every ticked Submit is one independent, complete call. No memory of the previous turn carries over. askLLM’s built-in debounce reuses a cached result only when the question, summary, persona, language, and system prompt are all identical to the last call. It never threads two questions into a conversation. The app’s own guidance says as much: untick Submit before editing your question, then re-tick it. Every edit can trigger a new call. This means your prompt must be self-contained. State the background, variables, and comparison you want in one shot. You cannot count on “I’ll clarify in the next message.”
八組改寫前／改寫後對照Eight before/after rewrites
| 改寫前 | 改寫後 | 補進了什麼 |
|---|---|---|
| 幫我分析這個資料 | 連續變項 simon_effect 在兩水準的類別變項 similarity（same／different）之間是否有差異？請建議該用哪個 jamovi 分析。 |
結果變項名稱與型態、分組變項名稱與水準數、設計（兩組比較）、要求的輸出（分析建議）。 |
| 這個資料有問題嗎？ | 變項 Household Income 的遺漏比例是多少？我該先用哪個分析檢查遺漏是否集中在 Political Preference 的特定水準？ |
具體變項名稱、想檢查的現象（遺漏是否集中在某個類別變項）、期望得到的下一步。 |
| 幫我做統計 | 連續變項 CalEstimate（估計攝取量）和實際攝取量之間有沒有關聯？該用哪個相關分析？ |
兩個變項的名稱與型態、關係類型（關聯而非差異）、要求的輸出。 |
| 這個結果好嗎？ | 對 simon_effect 跑完獨立樣本 t 檢定後，如果 Levene’s test 顯著，我該看 Welch 還是 Student 的結果列？ |
已完成的步驟（對哪個變項跑過 t 檢定）、判斷依據（Levene’s test 顯著與否）、想決定的具體問題。 |
| 幫我選模型 | 結果變項 mood.gain 是連續的，分組變項 drug 有三個水準（placebo／anxifree／joyzepam），該用哪個 ANOVA？ |
變項型態、水準數（三組）、明確要求（該用哪個 ANOVA）。 |
| 我要跑迴歸 | 用連續變項「每日使用時數」與類別變項 gender（兩水準）預測連續結果變項 WEMWBS_sum（心理幸福感量表總分），該用哪個迴歸分析？如何檢查預測變項之間的共線性？ |
預測變項與結果變項的名稱、型態、數量、分析目標（預測）、想要的額外檢查。 |
| 資料有離群值嗎？ | 變項 reaction_time 在 condition（congruent／incongruent）兩水準下的摘要顯示 max 遠大於 mean，我該用哪個分析畫圖檢查是否為離群值？ |
具體變項與分組、觀察到的線索（摘要中的 max 與 mean 差距）、想要的輸出（可執行的分析建議）。 |
| 這兩組一樣嗎？ | 類別變項 Gender（3 水準）在類別變項 Education（5 水準）之間的分布是否有關聯？該用哪個分析？ |
兩個變項皆為類別、各自水準數、關聯類型（列聯表關聯而非平均數比較）、明確要求分析建議。 |
| Before | After | What was added |
|---|---|---|
| Analyze this data for me | Does continuous variable simon_effect differ between the two levels of similarity (same/different)? Which jamovi analysis should I use? |
Outcome name and type, grouping variable and its levels, the design (two-group comparison), and the requested output. |
| Is something wrong with this data? | What is the missing rate for Household Income? Which analysis should I use first to see whether the missingness clusters by levels of Political Preference? |
The specific variable, the pattern being checked (whether missingness clusters by a categorical variable), and the desired next step. |
| Do the stats for me | Is there an association between continuous variable CalEstimate (estimated intake) and actual intake? Which correlation analysis fits? |
Both variable names and types, the relationship type (association, not difference), and the requested output. |
| Is this result good? | After running an independent-samples t-test on simon_effect, if Levene’s test is significant, should I read the Welch or Student row? |
The step already done (which variable the t-test was run on), the decision criterion (Levene’s test), and the specific choice to make. |
| Help me pick a model | mood.gain is continuous, drug has three levels (placebo/anxifree/joyzepam) — which ANOVA should I use? |
Variable types, the number of levels (three groups), and the explicit request (which ANOVA). |
| I want to run a regression | Predict the continuous outcome WEMWBS_sum from continuous daily screen time and categorical gender (two levels). Which regression should I use, and how do I check collinearity between the predictors? |
Predictor and outcome names, types, count, the goal (prediction), and the extra check requested. |
| Does the data have outliers? | The summary of reaction_time, split by the two levels of condition (congruent/incongruent), shows a max far above the mean — which analysis should I use to plot and check for outliers? |
The specific variable and grouping, the observed clue (the max-vs-mean gap in the summary), and the desired output (an actionable analysis suggestion). |
| Are these two groups the same? | Is there an association between the distribution of categorical Gender (3 levels) and categorical Education (5 levels)? Which analysis fits? |
Both variables are categorical, their level counts, the relationship type (a contingency association, not a mean comparison), and the explicit request for an analysis. |
動手做（5 分鐘）Try it (5 minutes)
- 挑一個你正想問 askLLM 的問題，先寫下目前腦中的版本（通常很模糊）。
- 對照上表，補進四件事：變項名稱、變項型態（連續／類別）、比較或關聯的方向、 你想要的具體輸出（例如「建議分析」或「該看哪個結果列」）。
- 把改寫後的一句話整段貼進 askLLM 的 Your question，勾選 Submit。
- 讀回覆，確認它回答的正是你改寫後的那句話，而不是原本模糊版本可能被誤解 的方向。
- 若還想追問，先取消 Submit 勾選，把追問內容也寫進同一個自足的問題裡， 再重新勾選——不要指望它記得上一次的內容。
- Pick a question you actually want to ask askLLM, and write down the vague version currently in your head.
- Using the table above, add four things: variable names, variable types (continuous/categorical), the direction of comparison or association, and the specific output you want (e.g. “suggest an analysis” or “which result row to read”).
- Paste the rewritten sentence into askLLM’s Your question box and tick Submit.
- Read the reply and confirm it answers exactly the rewritten question, not a possible misreading of the original vague one.
- If you want to follow up, untick Submit, fold the follow-up into the same self-contained question, and re-tick it. Do not expect it to remember the previous call.
帶走這句話One thing to remember
askLLM 每次 Submit 都是一次全新的獨立問答，沒有上一句可以依靠，所以你的問題 要在送出的那一刻就自己講完整個故事。
Every Submit to askLLM starts a brand-new, independent exchange with nothing carried over. So your question has to tell the whole story by itself, the moment you send it.
資料來源Data sources
上面八組對照用了以下四個外部素材，各自的授權標示於條目後。真實已發表研究標示為 「已發表研究」；lsj-book 的 clinicaltrial.csv 是教學用虛構資料，非真實研究， 標示為「教學資料集」。
- psyteachr Analysis（Mahrholz & Kuepper-Tetzel, 2025, CC BY 4.0）。第 1、4 組取自一項已發表研究（Zwaan et al., 2018 的 Simon task）， ch.7 Independent-samples t-test； 第 3 組取自一項已發表研究（Lopez et al., 2024 的湯碗研究）， ch.9 Correlation； 第 6 組取自一項已發表研究（Przybylski & Weinstein, 2017）， ch.11 Multiple regression （原始資料在 OSF：https://osf.io/bk7vw/）；第 8 組取自一項已發表研究 （Ballou et al., 2024）， ch.6 Chi-square （原始資料在 OSF：https://osf.io/6xkdg/）。CC BY 4.0 只需標示出處，不強制以相同授權分享。
- psyteachr Fundamentals of Quantitative Analysis(Bartlett & Toivo, 2024, CC BY-SA 4.0）。第 2 組取自一項已發表研究（Dawtry et al., 2015）， ch.11 Screening data。 CC BY-SA 4.0，本站以相同方式分享。
- psyteachr Data Skills（Nordmann, CC BY-SA 4.0）。第 7 組取自 Stroop 資料集。 CC BY-SA 4.0，本站以相同方式分享。
- 《Learning Statistics with jamovi》（Navarro & Foxcroft, 2025, Open Book Publishers, DOI 10.11647/OBP.0333, CC BY-SA 4.0）。第 5 組取自一份教學資料集
clinicaltrial.csv， ch.13 Comparing several means。 該章對事後檢定方法另有討論，想深入的讀者可自行前往閱讀；本頁不在此表態。 CC BY-SA 4.0，本站以相同方式分享。
The eight rewrites above draw on four external sources; the license for each is noted alongside it. A real published study is labelled “a published study”; the clinicaltrial.csv dataset from lsj-book is a teaching dataset, not real research, and is labelled as such.
- psyteachr Analysis (Mahrholz & Kuepper-Tetzel, 2025, CC BY 4.0). Rows 1 and 4 use data from a published study (Zwaan et al., 2018, Simon task), ch.7 Independent-samples t-test; row 3 uses data from a published study (Lopez et al., 2024, the soup bowl study), ch.9 Correlation; row 6 uses data from a published study (Przybylski & Weinstein, 2017), ch.11 Multiple regression (raw data on OSF: https://osf.io/bk7vw/); row 8 uses data from a published study (Ballou et al., 2024), ch.6 Chi-square (raw data on OSF: https://osf.io/6xkdg/). CC BY 4.0 requires attribution only, not share-alike.
- psyteachr Fundamentals of Quantitative Analysis (Bartlett & Toivo, 2024, CC BY-SA 4.0). Row 2 uses data from a published study (Dawtry et al., 2015), ch.11 Screening data. CC BY-SA 4.0; this page is shared under the same license.
- psyteachr Data Skills (Nordmann, CC BY-SA 4.0). Row 7 uses the Stroop dataset. CC BY-SA 4.0; this page is shared under the same license.
- Learning Statistics with jamovi (Navarro & Foxcroft, 2025, Open Book Publishers, DOI 10.11647/OBP.0333, CC BY-SA 4.0). Row 5 uses a teaching dataset from Learning Statistics with jamovi,
clinicaltrial.csv, ch.13 Comparing several means. That chapter also discusses post-hoc test choice; readers who want to go further can follow the link — this page takes no position on it. CC BY-SA 4.0; this page is shared under the same license.