Stop over-prompting reasoning models: 4 settings that matter
If your AI prompts still contain “double-check your work,” “think step by step,” or a wall of ALWAYS/NEVER rules, you’re writing for a model that no longer exists. Current reasoning models, including GPT-5.6, Opus 5, and Kimi-3, already run an internal verification pass and pace their own depth. Those extra lines don’t add safety. They add friction and burn tokens on work the model was going to do anyway.
What the data says
OpenAI found 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%. That’s not a marginal gain from cleaner writing. That’s a structural improvement from removing instructions the model was actively working around.
The four settings a model still can’t infer
The argument isn’t for shorter prompts as an end in themselves. A short but vague prompt still fails. The goal is a prompt where every remaining line does something the model’s defaults don’t already cover. According to the author, four settings qualify:
- Effort: Replace “think hard” or “think deeply” with the model’s official effort selector (low, medium, high, or max). Start low and raise it only when the task genuinely needs more depth. Note that the exact parameter name and accepted values differ across providers, so confirm before relying on it.
- Scope: State what the model should and shouldn’t touch. Without a scope boundary, a one-file fix can turn into a refactor of the whole module.
- Length: Specify a paragraph, a table, or three bullets. “Be concise” without specifics leaves the model guessing what to cut.
- Autonomy: The author recommends a three-level policy: act without asking on reversible low-risk steps, confirm before anything destructive or hard to undo, and for pure analysis or planning, inspect and report without touching code. Anthropic’s guidance on Opus 5 states directly that explicit verification instructions cause over-verification in modern models and should be removed.
What to delete from existing 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
- Duplicate rules: each constraint should appear exactly once
- Bare “be concise” with no specifics on what to keep or cut
What the replacement structure looks like
The author proposes a 2026 prompt template with six fields: Role, Objective, Success criteria, Constraints, Output format, and 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 that belong in every version of the prompt.
The article also includes a meta-prompt approach: feed your old prompt into a prompt that applies these rules and returns a trimmed version, so you don’t have to audit legacy prompts by hand.
One caveat: this approach targets reasoning models with explicit effort parameters and strong default self-verification. Older or smaller models without those defaults may still need the scaffolding this tip removes.
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.
