Stop prompting reasoning models like it’s 2023
If your prompts still say “double-check your work” or “think step by step,” you’re writing instructions for a model that no longer exists. That’s the argument from developer and educator Maxi Contieri in a practical breakdown of how to update prompts for current reasoning models.
What the research shows
OpenAI reports that trimming internal agent prompts, cutting repeated instructions, unnecessary examples, and irrelevant tool descriptions, raised evaluation scores by 10 to 15%. The same trimming dropped token consumption between 41 and 66%, and cut cost by up to 67%.
The reason: models like GPT-5.6, Opus 5, and Kimi-3 already run an internal verification pass and pace their reasoning depth against the task. Asking them to “verify” or “think deeply” doesn’t add behavior. It adds noise the model works around.
The four things a model still can’t infer
The argument isn’t “write short prompts.” A short but vague prompt still fails. The argument is to keep only the lines that cover ground the model’s defaults don’t already handle:
- Effort level: Use the model’s explicit effort selector (low/medium/high/max) instead of “think hard.” Start low and raise it only when the task needs more depth.
- Scope: State exactly what files, modules, or sections are in play. Without it, a one-file fix can turn into a full module refactor.
- Length: Name what to keep and what to cut rather than writing “be concise.”
- Autonomy policy: State when the model should act without asking, when it should confirm first, and when it should inspect and report only.
What to delete from old prompts
- Forced verification lines: “double-check your work,” “review before answering”
- Vague depth requests: “think deeply,” “think hard”
- ALWAYS/NEVER absolutes written for judgment calls rather than genuine invariants
- Bare “be concise” with no specifics on what to cut
- Duplicate rules inherited from older prompt versions
The replacement template
Contieri proposes a six-part structure for 2026 prompts: Role, Objective, Success criteria, Constraints, Output format, Stopping rules. Business rules, security limits, and data boundaries stay in regardless of how short the rest gets. Those aren’t scaffolding. They’re non-negotiable constraints.
One note on the effort selector: the parameter name and accepted values are model-specific. Confirm the exact syntax for the model you’re calling before relying on it.
Full breakdown with prompt examples at HackerNoon.
I’m founder of BizStack at Cagri Sarigoz LLC and a passionate advocate for entrepreneurs.
With over 14 years in tech, marketing, and AI, including my role as Head of SEO at CitizenShipper and co-founder of TaleBot at Intale AI, I’m dedicated to sharing genuine, useful product insights and tips.
At BizStack, I aim to cut through the digital noise to provide clear, actionable advice.
And more than all else, I’m a father to a (always) little girl and a husband.
Contact me at [email protected] for assistance.
