The Agent-Coding Failure Nobody Talks About: Right Code, Wrong Idea
gandazgul · reddit · 2026-09-18
The author argues the main agent-coding failure is no longer code quality but wrong ideas: after a half-formed description, the agent starts writing and 450 lines later the idea turns out wrong. Two "refuse-to-do" prompt modes fixed most of it:
- A mode that won't implement: it restates the idea first (catching half the ambiguity), then interviews one question at a time, only about choices that change the product. Batched question lists feel efficient but break because question 2 assumes an answer to question 1. Inspired by Matt Pocock's grill-me skill, which asks too many upfront details.
- A mode that won't edit: most usage is questions, not changes; a write-capable agent answering them yields a confident wrong answer plus three edited files. So: read-only, cites sources, says "no evidence" instead of filling gaps.
Both are packaged as skills, installable via npx skills@latest add gandazgul/runwield, MIT licensed.
More from coding & agent
- dingtalk-wiki-mcp: open-source MCP lets agents read and write DingTalk Wiki — modelcontextprotocol · 2026-09-18
- Models fill in blanks: a pre-execution gate that strips verdict authority from LLMs — Jay299792458 · 2026-09-18
- 6 prompts to turn research piles into finished content with Gemini Notebook and Claude — Aiden_Tech_Ai · 2026-09-18
- Step-by-Step Guide to Becoming an SRE: LLM Monitoring and Canary Deploys — ashishllm · 2026-09-18
- Catching AI-generated race conditions by simulating thousands of codebase timelines — cto_junior · 2026-09-18
- Hooking WeChat Up to Codex: Let AI Batch-Read Your Chats and Moments — sven_ai · 2026-09-18