"Be creative" prompts drift across agent models, checkable requirements don't
ClickOk5811 · reddit · 2026-10-04
A developer swapping the underlying model behind an agent's planning step found that "Be creative" in the reasoning prompt produced wildly varied, chaotic tool selection on one model but only a conservative nudge on another. The word does no fixed work: its effect is relative to each model's default behavior, so deterministic models read it as a bigger push while wander-prone models barely shift. What survives model swaps is anything stated as a checkable requirement — output schemas for tool calls, explicit tool eligibility constraints, defined fallbacks — because hard requirements aren't measured against a model's personal habits.
More from coding & agent
- Fortnox Doc MCP serves 377 API endpoints' docs straight into your AI assistant — modelcontextprotocol · 2026-10-04
- Teams are consolidating into monorepos to make them easier for AI agents to work with — neal_lathia · 2026-10-04
- Using Skills and Memory to audit spreadsheets with traceable findings and saved preferences — FellMentKE · 2026-10-04
- Project-based iteration keeps what landed in AI video, with Codex and Dreamina CLI automating the loop — FellMentKE · 2026-10-04
- Turning research notes into structured, editable proposals and reports with tone control for the audience — FellMentKE · 2026-10-04
- Global Macro Database update brings hundreds of years of macro data to Claude Code and Codex agents — SchoeneggerPhil · 2026-10-04