Which Old Prompting Habits Now Get in the Way

The strong models of 2026 recover intent better and do part of the reasoning on their own. So many techniques from the 2023 guides now don't help - they get in the way. You can't delete them blindly, but you can't keep them out of habit either: every technique has to earn its place through your eval, not through the authority of an old article.

Let's go through what most often needs changing.

• "You are a world-class expert" at the top of the request. This used to count as a warm-up; now it is weak cosmetics - the model already answers at its level. Instead of a title, describe the actual function, audience, data, decision, and boundaries of responsibility - the things that genuinely change behavior.
• "Think step by step, show your reasoning". Often unnecessary, sometimes harmful: you pay for a hidden chain of thought and get verbosity instead of an answer. Give the goal, the criteria, and the reasoning level you need, and ask for a verifiable result rather than thinking out loud.
• A giant prompt with every if/else. This is a brittle pseudo-program: one edit breaks a neighboring branch. Keep the minimally sufficient instructions and canonical examples, and move policy into code and narrow tools.
• JSON described in a paragraph. Unreliable - the model will keep breaking the format. Use structured output or a schema for tool parameters, then domain validation.
• The whole knowledge base in the context. Expensive, and it blurs attention. You need retrieval with filtering, metadata, citations, and an answerability gate - the RAG chapter is devoted to this.
• Multi-agent for every hard task. That is an architectural tax. Start with a single call or a deterministic workflow; separate agents by different contracts, tools, and isolated contexts, and parallelize only genuinely independent branches.
• Always a self-check and a critic agent. The effect is model-dependent: Anthropic explicitly warns about over-verification in Claude Opus 5. Check whether it actually helps in your case rather than adding it by default.
• Temperature 0 as a guarantee of determinism. A false guarantee: even at zero the system doesn't become predictable. Reliability comes from schema, validators, multiple runs, and a regression suite.

To catch these symptoms in your own text there is a simple local check - it looks for rituals, it doesn't pass sentence.

And above all - don't swap one cult for another.

Don't swap one cult for another. A short prompt isn't always better than a long one. The goal isn't the fewest words but the smallest set of high-signal information that reliably passes your evals.