AI Collaboration Report: Iteration 1 (2026-09-23 ~ 2026-10-09)

Version	Date	Changes
1.0	2026-10-07	Iteration 1
1.1	2026-10-09	Rewritten along the course guideline; full prompt log moved to a sub-page; the team's implementation notes added
1.2	2026-10-09	Development prompts in sections 2 and 5; prompt log rebuilt

Written by: Kim Jinho (PM) · Contributors: Kim Jinho, Kim Sooyeon, Park Junha Tools used: Kiro (agentic IDE on Claude models), Claude Full prompt log: AI Collaboration Report — Iteration 1 — Prompt Log

---

1. Where AI Was Used

Work	Tool	Where to check
Backend and app code to the team's specs: API wiring, fields, unit tests	Kiro, Claude	backend/apps/, APP/app/src/; e.g., e72d2ae
Figma frames updated to the team's change requests	Kiro	Figma board; prompt log, 1
Front-end screens from the Figma drafts	Claude	Branch front-iter1 (d72c176–0ab39af)
QA scripts and end-to-end tests	Claude	qa/demo/, APP/app/src/androidTest/
Review of the 0.9.0 code	AI reviewer (tool not recorded)	Section 4
Checks against live APIs; README and Wiki pages	Kiro	PR #1; Wiki c4bb132, fe8cf76, 3c2c9a5

Not used, on purpose: product scope (features, user stories, the Iteration 1 task list and priorities; section 4 shows why), the Figma drafts, merges into main, the final device checks and the demo video. The team made or approved every change on main and the Wiki. AI-generated code isn't marked in the code yet (section 7).

2. Prompt History (excerpt)

Kim Jinho → Kiro. The PM edited these to make the split of work explicit, so the wording differs from what was sent; the verbatim originals and outputs are in the prompt log above. Other members' notes: sections 4–5.

1. [09-23, two screenshots] Team sets the design, AI builds it: "…정한 건 이거야. 1. 버스 구간 색은 마을버스, 시내버스 등 종류에 맞게. … 피그마에서 UI를 참고해서 구현해줘. 수단별 색 매핑처럼 반복되는 구간도 같이. 더 바꿀 건 결과 보고 내가 정할게." → 804bb43, four new test files
2. [09-24] Team sets the rule, AI implements it: "도착 판별 규칙은 이렇게 정했어. 목적지 geofence 진입 + 2분 이상 체류 = 목적지 영역 도착 추정. 반경에 들어와도 2분을 못 채우고 나가면 도착이 아니야. 규칙은 이대로 가고, 너는 구현을 맡아줘." → ebd9569
3. [09-28, another agent's review pasted] AI checks, team decides: "…바로 고치지 말고, 항목마다 그 에이전트가 말한 것이랑 실제 코드랑 맞는지부터 하나씩 확인해줘. 그다음 먼저 고쳐야 할 우선순위를 정해봐. 개선 순서는 그거 보고 내가 정할게." → section 4
4. [09-30, Figma ⑬-d screenshot] AI looks for other sources: "…여기 "경사도"(1.2%, 0.8%, 0.6%)가 들어가 있는데, 이게 지금 우리 API에서 실제로 받을 수 없는 정보야. Kakao API 외에 다른 API에서 이 정보를 받을 수 있는 경로가 있는지 탐색해줘." → section 3

3. What AI Did Well

• Design checked against the APIs before building (prompt 4). Neither our API nor Kakao's bicycle route has elevation, so the slope figures on ⑬-d were mockups. AI listed three options, one of them computing slope from external elevation data, and left the choice to us; slope was dropped, not faked (Design 2.2.8).
• Specified, repetitive changes came with tests (prompts 1–2). The live route (e72d2ae) came with five new test files, app and server. CI passed 54 of 58 runs, and its contract check flagged a field the server didn't send yet (9a07472 → a21861d).
• Documents built from the code. The Testing page re-ran every suite before reporting (3c2c9a5); the Design schema came from the model definitions (fe8cf76).

4. Hallucinations and Errors

What the AI asserted or did	Why it was wrong	How we caught it	Cost
An implementation task breakdown added unrequested work and new features	Drifted from the app's intent and novelty	The member's review	Fixed by hand
Live route (0.7.0) drawn from the event's origin even after leaving	Meaningless once on the way	Team review of 0.7.0	New endpoint from the current position (e72d2ae)
Its refresh waited for 60 s and 150 m of movement	We asked for once a minute	Checked against the request	Condition removed (9619903)
The 0.9.0 review (prompt 3) called "the fastest route isn't the fastest across modes" a bug	Intended behavior	Team review, 10-08	Dropped from the known bugs (d20b620); its other four bugs were real (one fixed, 4362c51)

5. Prompt Revisions

Quoted from the originals, cut with "…".

• Live route. Before (09-25, map screenshot): "…가장 빠른 쪽도 표시해뒀으면 좋겠어. 이건 초록색으로 표시되게 하자." 0.7.0 drew it from the event's origin, even after leaving (7d90d90). After (same day): "…1분에 한번씩 유저의 위치를 파악해서 … 목적지까지 가장 빠른 경로를 다시 파악한 뒤, 지도에 표시하는걸로. … 보라색으로 하자. 점선 말고 그냥 선으로." (e72d2ae, 9619903) Naming the origin, refresh rule and line style left nothing to interpret.
• Arrival times at stops. Before (09-23): "…지하철은 #분#초뒤 도착, 버스는 #분#초뒤 도착." The app showed when to board. After (12 minutes later, with screenshots): "…이건 사용자가 8분56초 뒤에 탄다가 아니라, 5511버스가 정류장에 8분56초 뒤에 도착한다는 뜻이잖아." (b655552) Reading one number off the screenshot fixed what the number means.
• UI changes (team note). Before: changes described in text, without updating Figma, took several requests to match. After: show them as a picture, which pins down layout that words leave open.

6. Manual Fixes and Why

• Task breakdown: corrected by hand rather than re-prompted, because scope and novelty are the team's call.
• Facts only the team knew were edited on the Wiki directly: customers and competitors (Kim Sooyeon, f5c7ae9) and how this report describes the team's work (Kim Jinho, 2fde32c).
• Demo video: an AI-driven recording through adb was stopped after its first rehearsal needed fixes on several screens; the PM recorded the video by hand.

7. Takeaway for Iteration 2

• Write acceptance criteria first, ask AI for a plan limited to them, and show UI changes as a Figma frame or screenshot.
• Merge AI changes through pull requests with CI, mark AI-generated code (tool, date, reviewer), and have every member log their prompts verbatim, with agent hours and tokens per task.