Cease Over-Prompting: OpenAI’s New GPT-5.6 Pointers Change Every little thing – Decrypt

OpenAI printed a brand new prompting information for GPT-5.6 Sol, its newly launched flagship mannequin, and the principle message will really feel flawed to anybody who spent the final yr writing multi-page system prompts: cease writing a lot. The core thought is outcome-first prompting. Outline what attractiveness like, set the stopping circumstances, and get out of the way in which.

Detailed how-to directions, repeated type guidelines, examples that do not change habits—all of it’s now thought of noise.

OpenAI backs this with numbers: In inside coding agent assessments, leaner system prompts improved analysis scores by roughly 10–15% whereas chopping whole tokens by 41–66% and prices by 33–67%.

GPT-5 vs. GPT-5.6: What really modified

The GPT-5 prompting information, printed at launch in August 2025, was about including scaffolding. You bought XML persistence blocks telling the mannequin to maintain working till the issue was solved, detailed context-gathering templates that mapped precisely how one can parallelize searches and when to escalate, and power preamble scripts that narrated each step out loud.

The philosophy was calibrating eagerness—constructing express rails for when to go more durable or stand down.

GPT-5.6 principally would not want these rails. The brand new information tells you to trim: repeated guidelines, type directions that do not change habits, examples that do nothing, and course of steps the mannequin already handles reliably. So principally, that “ block with its parallel search batches and early-stop standards that used to assist is now scaffolding the mannequin has to parse round, not scaffolding that helps it.

What you really preserve is easier: the user-visible final result, success standards, stopping circumstances, and onerous constraints. The information’s mannequin of a very good immediate begins with “Resolve the shopper’s problem finish to finish”—then specifies precisely what completed appears to be like like, what actions to finish earlier than responding, and what to do when required proof is lacking. Not “be thorough.” Not “preserve going.” Simply: right here is the vacation spot.

The chance calculus additionally shifted. The information warns that GPT-5.6 follows immediate contracts intently, and that “conflicting guidelines can create extra instability than lacking element.”

An earlier mannequin would choose one instruction when it hit a battle. GPT-5.6 burns reasoning tokens making an attempt to reconcile each, which is slower, dearer, and infrequently flawed. In case your system immediate has overlapping guidelines—and most manufacturing prompts do—that is the factor to repair first.

Additionally OpenAI closely advises in opposition to utilizing the previous trick of resorting to absolutes like “at all times do that” or “by no means do this” to steer the AI’s habits in a selected route.

Two concrete additions spherical out the distinction. The primary is the textual content.verbosity parameter: As a result of GPT-5.6 is already extra concise by default than GPT-5.5, previous “be transient” directions now over-correct and make responses too brief. Set a worldwide default by way of the parameter, then override per process within the immediate. The second is a bit on Programmatic Instrument Calling—for bounded workflows the place code handles filtering, batching, or aggregating massive intermediate outputs and returns a compact consequence, offloading that work from the mannequin’s judgment solely.

However does it work?

We used the information to optimize our immediate for TYPE OR DIE, the first-person typing survival horror recreation we construct to benchmark a mannequin’s coding talents. The consequence was extra polished: GPT-5.6 Sol tackled the auto-aim logic extra effectively than on earlier runs, the visuals had extra coherence, and the general really feel of the sport was cleaner.

It took extra time to construct. The mannequin did not soar straight to code—it mapped the complete drawback first, deliberate every system earlier than writing a line. That is the information working as supposed. Outline the vacation spot; the mannequin chooses the route.

The brand new immediate is accessible on our Github so you may test it out.

You possibly can play the unique GPT 5.6 recreation by clicking on this hyperlink.

The sport created below the newer immediate, is accessible right here.

If you wish to push additional, or are too lazy to memorize all these new tips, you may construct your individual customized GPT and feed it the complete information as its information base. Configure it to research any immediate you throw at it, perceive the underlying logic, and rewrite it in GPT-5.6 type. You find yourself utilizing immediate engineering to engineer higher prompts.

Promptception. You are welcome.