Spec first, code second: one workshop experiment shows agents build the wrong app
Al_Grigor · x · 2026-10-29
The author ran a comparison in his workshop: a one-shot prompt — "implement a tool for weekly project feedback" — made the agent ask zero questions and build a CLI for tracking project health. The same idea, scoped by having ChatGPT interview the author one question at a time, produced a team retrospective tool with anonymous feedback, clustered cards, voting, and action items.
His recommended spec covers: users, workflow, MVP, out-of-scope, and decisions the agent can't make. The core lesson: technically fine code can still miss the product — any decision left unstated gets made by the agent.
Workshop recording linked in the post.
More from coding & agent
- Devv, early AI search engine serving 1M+ developers, shuts down to build agent workspace Multica — jiayuan_jy · 2026-10-29
- Aristotle agent stalls for ages on a 200-line Verilog file with no progress signal — thomasahle · 2026-10-29
- Four-tier human takeover architecture for low-latency voice LLM pipelines — Nick-Can9972 · 2026-10-29
- Lightpanda, a Zig-built browser for AI agents, uses 16× less memory than headless Chrome — techNmak · 2026-10-29
- How OpenClaw (9K stars day one) and Instinct ($10B, 14 employees) defined personal agents — sujingshen · 2026-10-29
- VoidZero, now part of Cloudflare, open-sources Void: full-stack Vite deploy platform — cnakazawa · 2026-10-29