+ Marks uncertainty and captures all applicable constraints.
- Some owners are inferred, and success criteria remain vague.
Records why each decision was made, so the next session does not reopen it.
| Category | Using AI › Context management |
|---|---|
| Tags | SummarizingDraftingTemplate |
Write a handoff so this work can continue in a fresh conversation. **The reader knows nothing. Assume no shared history.** **Include:** 1. **The goal** — one sentence. What "done" means 2. **Decisions already made, and why.** ***The reason matters more than the decision*** — **without it the next session re-opens settled questions** 3. **What was tried and did not work**, with the reason. *Otherwise it gets tried again* 4. **Current state** — what exists now, where it is 5. **The immediate next step**, concretely 6. **Constraints that still apply** — and ***which ones are decisions versus which are hard limits*** 7. **What is still open**, and who decides it **Rules:** - ***Do not include the history of how we got here.*** **Only what changes what happens next** - **Write terms so they cannot be misread.** A name only meaningful in the old conversation gets defined - ***Mark anything uncertain as uncertain.*** **A handoff that states a guess as a decision is worse than one that leaves a blank** - Keep it to one screen. **If it will not fit, that is a sign the work needs splitting**
The field most often missing from a handoff is the reasoning — without it, settled questions get reopened and failed attempts get retried. This carries only what changes the next action.
ChatGPT best handles uncertainty and hard limits. Gemini is well structured but presents unsupported assumptions as facts, while [C] is missing.
+ Marks uncertainty and captures all applicable constraints.
- Some owners are inferred, and success criteria remain vague.
+ Clearly separates decisions, rationale, and next actions.
- Invents uptime, TTL, ownership, and deployment details.
| Criterion | ChatGPT | Gemini | Leader |
|---|---|---|---|
| Instruction following | 9 | 7 | ChatGPT +29% |
| Accuracy | 9 | 5 | ChatGPT +80% |
| Specificity | 9 | 8 | ChatGPT +13% |
| Structure | 9 | 9 | Tie |
| Right length | 9 | 8 | ChatGPT +13% |
Scored 1–10 by gpt-5.6-sol with model names hidden (2026-09-25). This is an AI review, not a measurement.
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.
The work so far: 쇼핑몰 '오늘장터'의 장바구니 API 리팩터링 작업. 기존엔 세션 기반 장바구니였는데 비회원 장바구니가 유지되지 않아 DB 기반으로 전환하기로 결정(비회원 이탈률 18% 감소가 목표). cart_items 테이블을 새로 설계해 마이그레이션 스크립트까지 작성 완료. 기존 세션 기반 로직은 삭제하지 않고 feature flag로 병행 운영 중(안전한 롤백을 위해). Redis 캐싱을 붙이려다 캐시 무효화 로직이 복잡해져서 보류함. What comes next: cart_items 마이그레이션을 스테이징 환경에 적용하고, feature flag를 5% 트래픽에 먼저 켜서 에러율을 모니터링하기.
목표
비회원 장바구니가 유지되도록 DB 기반 API로 전환하고, 비회원 이탈률을 18% 낮추는 것을 목표로 한다.
확정된 결정과 이유
cart_items 테이블을 도입한다. 세션 기반 장바구니는 비회원 상태에서 유지되지 않기 때문이다.cart_items 테이블 설계 완료.스테이징 환경에 cart_items 마이그레이션을 적용한 뒤, DB 기반 로직 feature flag를 전체 트래픽의 5%에 활성화하고 에러율을 모니터링한다. 이상이 있으면 세션 기반 로직으로 롤백한다.
| Chat Summary and Export Instructions | |
| Context Migration | |
| Continue and Recap Assistant | |
| Plan how to split long material for a model | |
| Find instructions that contradict each other |