Four spec-driven coding tools tested on one real feature: 13, 24, 29 tasks and 84 Allow clicks
SSShken · reddit · 2026-09-19
After criticism that his earlier comparison only measured onboarding, the author built a real TypeScript/Express/SQLite/React project with plain Claude Code, gave identical copies to OpenSpec, Spec Kit, BMAD and Kiro, and asked each to add the same one-line feature — with a planted trap: the server already supported it.
- All four shipped correctly and none touched the server or invented endpoints; all read the code before planning
- They diverged elsewhere: task lists of 13, 24, 29, and none; one asked before adding 3 dev dependencies, another added 7 silently; two left test data in the database; one committed straight to main; one required 84 manual Allow clicks and most of a free credit tier
- Everything counted by hand, every diff read manually; the missing control run (plain Claude Code, no workflow) comes next
Related event: Same prompt, four spec-driven coding tools: 4 minutes to 2 hours(2 posts)→
More from coding & agent
- Why LLM-Written Docs Fail: Models Optimize for Punchiness, Not Readability — mervenoyann · 2026-11-04
- OpenAI DevDay demo shows forking agent task history at any step — paw_lean · 2026-11-04
- Veteran game dev: "AI builds a full game in a day" posts are mostly hype — bendee983 · 2026-11-04
- Dev builds interactive Excel-to-dashboard with ChatGPT, shares the prompt — TermRemarkable665 · 2026-11-04
- Agents cut dev time from weeks to a day; bottleneck moved to specs and review — Suspicious_Orchid770 · 2026-11-04
- Orgo, a YC-backed platform giving AI agents their own cloud desktops, claims 1,000+ teams — Scobleizer · 2026-11-04