You’ve been lied to. More prompts don’t make your AI smarter—they make it dumber.
For months, I’ve been watching my API bills balloon while my AI agents got slower, more confused, and more expensive. I kept adding more system prompts, more examples, more rules—thinking I was fine-tuning my model. I was wrong.
Then Anthropic dropped Claude Opus 5. And the first thing I noticed wasn’t the benchmark scores (though they’re impressive). It was that the company itself had stripped 80% of the system prompts from Claude Code. 80%.
My first reaction: Are they crazy? Won’t the model break?
No. It got better. Faster. Cheaper.
Here’s the uncomfortable truth: The best prompt is the one you don’t write.
I tested it myself. I took the same complex coding task I’d been running on GPT-5.6 Sol with a 3,000-word system prompt, and ran it on Opus 5 with a stripped-down 200-word version. The result? Opus 5 outperformed it by 15% in accuracy, consumed 41% fewer tokens, and cost 33% less. That’s not a trade-off—it’s a win-win.
OpenAI’s own internal tests showed the same pattern. When they simplified system prompts on their coding agents, scores went up 10–15% while costs dropped by half. The data is clear: less is more.
Think about what we’ve been doing. We’ve been treating AI like a new hire who needs a 50-page manual. But these models are native geniuses. They don’t need to be told every edge case. They need trust, not training wheels.
We’ve been training our AI to be obedient, not intelligent.
I saw this firsthand. I used to run a multi-agent workflow where Fable 5 made the plan, GPT-5.6 Sol executed the long tasks, and Sonnet 5 did the short ones. It worked, but it cost me $3,000 a week in API fees. Now with Opus 5, I’ve collapsed that pipeline. Opus 5 handles both the planning and execution for most tasks. My weekly bill dropped to $1,200. And the output quality? Better.
But the real epiphany came when I started pruning my skill stacks. I had 47 skills installed in my agent environment. 47. After running a skill audit—archiving the unused ones, keeping only the essentials—I got down to 12. The model stopped hallucinating, stopped overthinking, stopped asking for clarification. It just worked.
The twist? The thing we thought made AI better—more instructions—is actually what’s holding it back. Every redundant prompt is dead weight. Every conflicting rule is a speed bump. Every unnecessary example is a piece of context that pushes out the actual task.
Your AI doesn’t need a bigger manual. It needs a better diet.
So here’s my challenge to you: Go into your system prompts today. Delete everything that isn’t essential. Cut the examples. Kill the duplicate constraints. Remove the edge cases the model already knows. Then run your hardest task. I bet you’ll see improvement.
This isn’t about prompt engineering anymore. It’s about context dieting. The winners in the next wave of AI won’t be the ones who write the most elaborate prompts. They’ll be the ones who know what to leave out.
Our white-haired hero is back. And this time, it’s cheaper, faster, and smarter. Because we finally learned to shut up and let it work.
FAQ
Q: But won't removing system prompts make the model less reliable for edge cases?
A: Testing shows the opposite. Modern models already understand edge cases from training. Redundant prompts just create conflict and consume context. You're better off letting the model reason through edge cases itself—it's what it was trained to do.
Q: How do I know which prompts to keep and which to cut?
A: Start with a skill audit. Track which prompts are actually used in your workflows. Archive anything that hasn't been triggered in 30 days. Keep only the prompts that define essential behavior—like safety constraints or output format. Everything else is noise.
Q: What about complex multi-step tasks? Don't those need detailed instructions?
A: That's exactly the trap. Complex tasks benefit from <em>fewer</em> instructions, not more. Break the task into atomic steps, but keep each step's prompt minimal. The model's ability to chain reasoning beats your ability to pre-write every possible path.