"Be careful" isn't an instruction — it's a decision you're offloading to the model
ClickOk5811 · reddit · 2026-09-03
The author recalls a planning meeting where someone proposed fixing inconsistent AI output by "just telling it to be thorough." Their point: that's not an instruction — thorough has no baseline, no specific risks, no acceptance bar, so the model silently picks an interpretation on your behalf.
Key arguments:
- "Be thorough," "act as a senior engineer," "use good judgment" are the same failure in different clothes: they relocate a decision instead of making it.
- What works is acceptance-criteria-style prompts: not "handle errors properly" but "return this specific error shape on validation failure, log this specific context, do not retry on validation failures" — nothing left to interpret.
- A well-specified instruction and a well-written acceptance criterion are nearly the same artifact; most teams already know how to write one but haven't connected it to the other.
More from coding & agent
- Magnitude open-sources a local-model inference server that plugs into Claude Code, Codex, Cline and more — magnitudedev · 2026-09-03
- Open-source ai-knowledge-graph turns unstructured text into interactive knowledge graphs via LLMs — tom_doerr · 2026-09-03
- Simulated GitHub run by agents shows AI-agent collaboration follows its own scaling laws — tongshuangwu · 2026-09-03
- Open Notebook: a 30k-star local NotebookLM alternative with 4-speaker podcasts — JafarNajafov · 2026-09-03
- AFK Pilot Cloud launches: run your AI coding agents on a personal cloud machine for $5/mo — PawelHuryn · 2026-09-03
- Context engineering over prompt hacks: a 4C framework for briefing LLMs like new hires — bigaiguy · 2026-09-03