AI Code Builder — Methodology & Cycle Framework
Research Methodology & Cycle Framework
Karena inti project ini adalah riset AI (bukan pure engineering build), tim bekerja dengan pendekatan research cycle, bukan sekadar sprint fitur. Setiap cycle punya hipotesis yang diuji, kriteria keluar (exit criteria) yang jelas, dan keputusan eksplisit di akhir: lanjut, revisi, atau stop.
Back to: [AI Code Builder — Project README](Research: AI Code Builder)
2.1 Kerangka Cycle: Discover → Hypothesize → Experiment → Measure → Iterate (DHEMI)
flowchart LR
A["Discover
(Riset literatur, benchmark,
kumpulkan data awal)"] --> B["Hypothesize
(Rumuskan hipotesis &
target metrik)"]
B --> C["Experiment
(Build model/prototype,
generate code)"]
C --> D["Measure
(Testing harness,
scoring vs threshold)"]
D --> E{"Capai Exit
Criteria?"}
E -->|Ya| F["Iterate →
Lanjut ke Cycle berikutnya"]
E -->|Tidak| G["Revisi Hipotesis /
Pivot Pendekatan"]
G --> B- Discover — riset literatur & benchmark (Design2Code, Pix2Code, dsb), kumpulkan sample data awal.
- Hypothesize — rumuskan hipotesis konkret, misal: "Classifier berbasis vision-language model bisa mendeteksi feature Auth dengan F1 ≥ 0.85 dari 100 sample screenshot."
- Experiment — bangun model/prototype/code generator sesuai hipotesis.
- Measure — jalankan testing harness, ukur terhadap metrik & threshold (lihat [Metrics](Success Metrics) & [Thresholds](Minimum Score / Threshold)).
- Iterate — jika exit criteria tercapai, cycle dianggap selesai dan tim lanjut ke cycle berikutnya; jika tidak, kembali ke tahap Hypothesize dengan revisi (bukan mengulang dari nol).
2.2 Definisi Cycle di Project Ini
| Cycle | Nama | Fokus Utama | Exit Criteria (ringkas) |
|---|---|---|---|
| Cycle 0 | Discovery & Baseline | Riset literatur, kumpulkan dataset awal, tentukan metrik & threshold | Dataset awal siap, metrik & threshold disepakati tim |
| Cycle 1 | Core Detection & Generation (Auth, Frontend-first) | Classifier Auth + code generator dasar untuk 1 platform prioritas | Classifier F1 ≥ 0.85, code hasil generate compilable ≥ 95% |
| Cycle 2 | Testing Loop & Feedback | Bangun testing harness penuh + feedback loop retraining | Semua metrik soft-gate tercapai minimum untuk platform prioritas |
| Cycle 3 | Multi-Platform Expansion | Perluas ke Mobile & Backend, reuse shared Feature Detection | Ketiga sub-project punya pipeline generate + testing dasar berjalan |
| Cycle 4 | Optimization & Hardening | Optimasi latency, otomasi retraining, kesiapan GA | Semua metrik capai target ideal di ketiga sub-project |
Setiap Task di breakdown per sub-project akan diberi label Cycle supaya tim tahu di iterasi mana task tersebut relevan dikerjakan — bukan hanya urutan prioritas P0/P1/P2, tapi juga kapan secara siklus riset.
2.3 Ritual per Cycle (ala lab research)
| Ritual | Kapan | Tujuan |
|---|---|---|
| Cycle Kickoff | Awal cycle | Menyepakati hipotesis, metrik target, dan exit criteria bersama tim |
| Weekly Experiment Review | Mingguan selama cycle berjalan | Review progres eksperimen, blocker, dan preliminary result |
| Cycle Retrospective | Akhir cycle | Evaluasi exit criteria tercapai/tidak, catat learning ke [Experiment Log]([[ai-code-builder/experiment-log |
| Go/No-Go Decision | Akhir cycle | Keputusan eksplisit: lanjut ke cycle berikutnya, revisi hipotesis, atau stop/deprioritize |
2.4 Reporting Berkala
Setiap project wajib menghasilkan 2 jenis report berkala untuk transparansi progress dan traceability keputusan:
Weekly Report (per cycle iteration)
- Kapan: tiap akhir minggu dalam cycle
- Format:
YYYY-cycle-N-wW.md(contoh:2026-cycle-1-w3.md) - Isi: hypothesis minggu ini, tasks completed/in-progress, blockers, metrics vs threshold, decision (Lanjut/Revisi/Stop), next week focus
- Template: [Weekly Report Template](Weekly Report Template)
Monthly Report
- Kapan: tiap akhir bulan
- Format:
YYYY-MM.md(contoh:2026-08.md) - Isi: executive summary, highlights/lowlights, cycle status, metrics trend, decisions summary, risk updates, up next month
- Template: [Monthly Report Template](Monthly Report Template)
Cara Pakai
- Copy template yang relevan, rename sesuai convention.
- Isi field kosong — link ke task, metric, dan risk di [Risk Register](Risk Register) pakai wikilink.
- Commit ke repo
notes/ai-code-builder/reports/{weekly|monthly}/. - Update [Reports Index](Reports Index) supaya report baru muncul di daftar.
Lihat juga
- [Reports Index](Reports Index) — daftar semua report
- [Experiment Log](Experiment Log Template) — entry experiment yang di-link dari weekly report