Type: url
Original: https://techblog.sakurug.co.jp/article/jdqkqgjavou/
Status: generated
Language: ja
Summary
An engineer at SAKURUG reports a small experiment in which Claude Code built the same customer-management web app under five cumulative harness conditions — prompt only, CLAUDE.md, a PostToolUse hook running lint/typecheck/test, a UI checklist plus a Stop completion gate, and a “Gate TDD” stop gate split into Red/Green/Refactor stages. All five final builds passed lint, typecheck, unit and e2e checks, so the differences showed up instead in how often the hooks and gates blocked the agent mid-run and in what the agent was allowed to count as finished; a rerun of the prompt-only condition produced a markedly different UI language, seed data and file layout. The author concludes that a harness raises the floor and fixes the completion criteria rather than raising the best-case output, and proposes an adoption order for individual developers while noting the small number of runs.