Your Model Is Fine, Your System Prompt Is Sabotaging You

The most shocking signal is that OpenAI’s Codex system prompt contains absurdly prescriptive design rules—border radii, icon libraries, forbidden empty states—that make even powerful models like 5–6 produce generic “slop.” These rules waste tokens and were only partially removed after a YouTuber raised hell. The real bottleneck isn’t model intelligence; it’s the agent harness.

Claude Code’s workflow feature lets models define programmatic sub-agent orchestration in code, which is bounded and token-efficient. Using OpenAI’s 5–6 inside Claude Code yielded quarter the token usage of Codex’s Ultra mode and better output quality, because workflows end and don’t spiral.

The tradeoff: moving to Claude Code for orchestration means juggling separate tools and subscriptions, and model-specific formatting quirks still appear. But for teams tackling complex multi-step coding tasks, this approach reduces cost and raises consistency.

Engineering leaders should audit the system prompts and agent frameworks their teams rely on. A model’s benchmark means little if the harness steers it toward wasteful, rigid behavior. Expect a shift in the AI tooling battle from model specs to prompt engineering and workflow design.