AI Collaboration Report – Iteration 1

Document Revision History

Version	Date	Author	Changes
0.1	2026-10-06	Dongkun Moon	Created the page with the Iteration 1 report for body analysis (P9, P10).
0.1	2026-10-06	KiHwan Kim	Add content for infrastructure & gitops deployment (P6, P7, P15).
0.1	2026-10-07	Hyeon U Jeong	Added the Iteration 1 wardrobe report: AI-assisted development, Korean prompts, human corrections and reflection.
0.1	2026-10-09	Juhyeon Choi	Added the Iteration 1 report for project planning, documentation, wireframes, and iteration coordination.

---

1. Where AI Was Used

Claude Code (Claude Opus 5.5) supported body-measurement logic, Android screens, debugging, and tests (backend/body_analysis/pose.py; PR #12, PR #15, PR #17). Codex supported wardrobe capture, landmark/ARCore geometry, Gemini analysis, editable results, and persistence (PR #18, PR #20, PR #21); it also generated k3s/Argo CD manifests and CI configuration (PR #2, PR #4, PR #5). ChatGPT supported the proposal, Requirements & Specifications, Design Documentation, wireframes, user flow, and task schedule. Humans retained scope and task ownership, real-device verification, cluster setup, and final decisions. AI-generated code headers are identified in PR #23.

2. Prompt History

• Dongkun (10-01): “Context: a teammate reports that analysis keeps failing for some gallery photos, apparently depending on the image's pixel size. Task: reproduce it, find the root cause and fix it. Output: the cause, the fix and evidence that it works.”
• Kihwan (09-24): “서버 deploy 전에 main에서 검증하는 CI 짜줘”
• Juhyeon (09-23): “지금까지 작성한 StyleMate proposal을 전체적으로 검토해주세요. Target Customers, Motivation, Innovation이 서로 잘 연결되는지 보고, 우리 서비스의 특징이 더 잘 드러나도록 수정하려면 어떻게 하면 될지 알려주세요.”
• Hyeon U (10-03): “기존 테스트용 local-dev를 제거하고, develop 브랜치의 실행 방식과 FE·BE 구조에 옷장 기능을 통합해주세요.”

3. What AI Did Well

• Body analysis: MediaPipe input-width padding and a regression test fixed a crash in backend/body_analysis/pose.py (L81–92) (b0b1186; PR #15), about 20 minutes after the report.
• Wardrobe: Codex connected capture → analysis → correction → save across analysis.py, editor.py, views.py, and GarmentEditor.kt (1d342c1, d96049c).
• Infrastructure: Manifests followed the human-chosen directory and namespace design (3260866, PR #2).
• Documentation: ChatGPT helped reveal missing requirements and navigation states; no time-saved figure was recorded.

4. Hallucinations / Errors

• Body analysis: The agent incorrectly claimed that PR #16 would automatically retarget after PR #15 merged. A later repository check revealed the mistake, requiring PR #17 and an additional CI run.
• Infrastructure: Codex added unnecessary API endpoint checks before the container was ready. Repeated CI failures on main exposed the problem, requiring two follow-up PRs (PR #4, PR #5).
• Wardrobe: AI suggested models without confirming deployable weights and runtime compatibility. Further investigation revealed that KGDet could not replace HRNet, resulting in additional evaluation effort.
• Documentation: ChatGPT sometimes described planned features as implemented. Juhyeon caught these inconsistencies by checking the actual implementation and consulting feature owners, requiring manual verification and corrections.

5. Prompt Revisions

Kihwan (CI validation scope)

• Before: “서버 deploy 전에 main에서 검증하는 CI 짜줘”
• After: “CI는 컨테이너 환경에서 빌드했을때 서버가 정상적으로 healthy하게 뜨는지만 확인 후 정리하면 충분해”

Why it worked: The revised prompt narrowed the validation scope, removing unnecessary endpoint checks that had caused repeated CI failures (PR #4, PR #5).

Dongkun (measurement accuracy display)

• Before: “Show the measurement accuracy as a percentage.”
• After: “Base it on the benchmark bodies' known measurements: show a per-measurement accuracy % only when one of our benchmark photos is analyzed.”

Why it worked: The revised prompt specified the ground truth and restricted accuracy display to benchmark photos, preventing unsupported accuracy claims.

6. Manual Fixes and Why

Kihwan manually installed/configured k3s and Argo CD when agent instructions confused his local machine with Azure VM. Hyeon U provided real garment/phone input and judged the UI and feature scope; the recorded code corrections were still implemented through Codex follow-up prompts. Dongkun manually reviewed screens and kept product decisions such as requiring both photos and restricting accuracy labels to benchmark images; he reported no hand-edited code. Juhyeon decided scope with the team and checked documentation against actual implementation and feature owners.

Next Iteration

Test with real input early, specify constraints and ground truth, and distinguish planned functions from implemented work. See the linked Wiki Iteration 1 – Prompt Log for the full prompt history.