You’ve spent the last two years hoarding prompt engineering tricks. Role-playing setups, step-by-step logic, “take a deep breath and think carefully.” You built a fortress of instructions trying to force the AI to do what you want.
What if I told you that fortress is now crushing you?
Your meticulously crafted prompts aren’t a safety net; they’re technical debt.
Recently, Boris Cherny, the creator of Claude Code, revealed that when adapting to the new Opus 5 model, they deleted over 80% of their system prompts. The bulk of those rules were patches for old model incompetencies. When they stripped them away, the AI actually got smarter.
The hard truth is this: the AI evolved, but your prompt engineering habits didn’t.
We used to have to micromanage early models because they were dumb. If it missed a step, you added a rule. If it hallucinated, you added a boundary. Over time, your prompts became bloated graveyards of past failures.
But today’s models—like GPT-6 Astra—are hyper-sensitive to instructions. They don’t just ignore your outdated rules; they diligently execute them.
Stop teaching the AI how to do its job. It already knows how better than you do.
When you leave conflicting or overly broad rules in your prompt, modern AI will literally stop in its tracks to ask for clarification, or trigger massive, unnecessary testing protocols to satisfy your paranoia. Researchers at ETH Zurich and LogicStar.ai tested this. They found that adding bloated context files didn’t improve coding agents’ success rates. It just increased compute costs by 20%.
You are paying a 20% premium for the illusion of control.
Eric Provencher, who leads dev experience at OpenAI, agrees: prompts designed for older models are now actively suffocating the AI’s ability to find efficient paths. You’re locking a genius in a room and forcing it to follow an outdated manual.
So, what do you keep, and what do you kill?
Delete the “how.” Kill the rigid, step-by-step instructions. Delete the emotional encouragement (“You are a world-class expert…”). Delete the global mandates like “run all tests regardless of change size.”
Keep the “what” and the “what not to.” Keep the information the AI cannot guess—internal data, project history, naming conventions. Keep the choices that only humans can make—goals, priorities, risk tolerance. Keep the hard boundaries—security, permissions, irreversible actions.
Use code for deterministic checks (like JSON validation) and use natural language for intent. The clearer the division, the more stable your system.
Effective prompt engineering isn’t about addition anymore; it’s about ruthless subtraction. Start with a blank slate. Run a test. Add an instruction back only when the AI repeatedly fails at a specific task. If a rule doesn’t improve the result, kill it.
Length is an illusion; clarity is the currency.
The smarter the model gets, the more you need to resist the urge to dictate every step. Your job isn’t to be a helicopter parent. Your job is to define the goal, set the boundaries, and get out of the way.
FAQ
Q: If I delete my step-by-step instructions, won't the AI just skip steps and make mistakes?
A: If you're using a modern model like Opus 5 or GPT-6 Astra, no. These models already understand how to sequence tasks. If you remove a rule and it fails, add it back. But don't preload your prompt with paranoid instructions for failures that haven't happened yet.
Q: What is the actual financial impact of bloated prompts?
A: Beyond the wasted time of reading useless context, bloated prompts trigger unnecessary actions. ETH Zurich research shows that overly prescriptive context files increase compute costs by 20% or more because the AI diligently executes your unnecessary rules—like running full test suites for minor edits.
Q: Is traditional prompt engineering completely dead?
A: Micromanagement is dead. The future isn't 'prompt engineering,' it's 'context architecture.' You no longer dictate the path; you define the goal, supply the necessary internal data, set hard boundaries for safety, and let the AI figure out the most efficient path.