48/ 50
＋ Precisely follows the format and stays concise.
－ The medium confidence for publishing seems too low.
Pull only the action items out of a meeting
Extracts commitments with a confidence rating, resolves relative deadlines to dates, and never assigns an owner the record does not name.
| Category | Office work › Meetings |
|---|---|
| Tags | AnalyzingOffice workerTable |
Extract only the action items from this meeting record. **Do not summarize the discussion.**
Output a table: action / owner / deadline / the statement it rests on / confidence.
Rules:
1. Capture soft statements — "we should probably", "let us take a look" — as candidates, marked **confidence: low.**
2. Where no owner was named, mark it unassigned. **Do not assign one.**
***An acknowledgment is not an assignment.*** Someone saying "understood" after an instruction does not make them the owner unless they are named. **Where this is genuinely unclear, mark it unassigned and note the exchange.**
3. Where a deadline is relative ("next week"), compute the date from the meeting date and give it in parentheses alongside the original wording.
4. Below the table, collect the items that **need an owner confirmed.**
5. Exclude anything mentioned as already done.
6. Where two statements produce conflicting deadlines for the same action, list both rather than choosing.After pasting, fill in the fields at the bottom (Meeting date · Meeting notes)
What this prompt does
Soft commitments are where actions get lost. This captures them as low-confidence candidates rather than dropping them, and collects the ones needing an owner.
Model comparison
ChatGPT has the best balance of format, accuracy, and length. Gemini misreads a deadline conflict, while Claude is specific but exceeds the action-only scope.
42/ 50
＋ Clearly explains unassigned owners and context.
－ It wrongly labels compatible timing as a conflict.
39/ 50
＋ Carefully distinguishes assignment from acknowledgment.
－ It adds excess material and misses a low-confidence rule.
| Criterion | ChatGPT | Claude | Gemini | Leader |
|---|---|---|---|---|
| Instruction following | 10 | 7 | 9 | ChatGPT +11% |
| Accuracy | 9 | 8 | 7 | ChatGPT +13% |
| Specificity | 9 | 10 | 9 | Claude +11% |
| Structure | 10 | 8 | 9 | ChatGPT +11% |
| Right length | 10 | 6 | 8 | ChatGPT +25% |
Scored 1–10 by gpt-5.6-sol with model names hidden (2026-09-24). This is an AI review, not a measurement.
Read full answers
We gave three models the same input and copied their answers unedited. Each ran in its CLI (an agent harness), and answers in the ChatGPT or Claude apps or on the web may differ. Outputs are in Korean.
Meeting date: 2026-09-22 Meeting notes: [주간 기획 회의 / 참석: 김부장, 이과장, 박대리, 최사원] 김부장: 지난주 얘기한 랜딩페이지 개편, 어떻게 됐나요. 이과장: 시안은 나왔는데 카피가 아직입니다. 김부장: 카피는 마케팅에서 받기로 한 거 아니었나. 박대리: 마케팅 쪽에 요청은 넣었는데 답이 없어서요. 김부장: 그럼 다시 한번 찔러보죠. 그리고 이번 달 안에는 올려야 합니다. 이과장: 네 알겠습니다. 최사원: 저 그리고 문의 폼 에러 나는 거 아직 안 고쳐졌는데요. 김부장: 아 그거. 한번 봐야겠네요. 박대리: 제가 볼게요. 김부장: 분석 툴도 붙여야 하는데... 그건 좀 나중에 하죠. 이과장: 개편 끝나고 붙이는 게 맞을 것 같습니다. 김부장: 그렇게 갑시다. 오늘은 여기까지.
| 액션 | 담당자 | 기한 | 근거가 되는 발언 | 확신도 |
|---|---|---|---|---|
| 마케팅에 카피를 다시 요청하기 | 미지정 | 없음 | “그럼 다시 한번 찔러보죠.” 이과장은 “네 알겠습니다”라고 했으나, 담당자 지정은 아님 | 낮음 |
| 랜딩페이지 개편본 게시하기 | 미지정 | 이번 달 안(2026-09-30) | “그리고 이번 달 안에는 올려야 합니다.” | 중간 |
| 문의 폼 오류 확인하기 | 박대리 | 없음 | 박대리: “제가 볼게요.” | 높음 |
| 분석 툴 연동하기 | 미지정 | 개편 완료 후 | 김부장: “분석 툴도 붙여야 하는데... 그건 좀 나중에 하죠.” 이과장: “개편 끝나고 붙이는 게 맞을 것 같습니다.” | 낮음 |
담당자 확인이 필요한 항목:
- 마케팅 카피 재요청 담당자
- 랜딩페이지 개편본 게시 담당자
- 분석 툴 연동 담당자
More in this category
| Brief someone who missed the meeting | |
| Write the follow-up email after a meeting | |
| Turn minutes into a chat message | |
| Split a meeting outcome by recipient | |
| Créer une Carte Mentale pour Séance d'Idéation |